CCSI: Continual Class-Specific Impression for data-free class incremental learning.

In real-world clinical settings, traditional deep learning-based classification methods struggle with diagnosing newly introduced disease types because they require samples from all disease classes for offline training. Class incremental learning offers a promising solution by adapting a deep network trained on specific disease classes to handle new diseases. However, catastrophic forgetting occurs, decreasing the performance of earlier classes when adapting the model to new data. Prior proposed methodologies to overcome this require perpetual storage of previous samples, posing potential practical concerns regarding privacy and storage regulations in healthcare. To this end, we propose a novel data-free class incremental learning framework that utilizes data synthesis on learned classes instead of data storage from previous classes. Our key contributions include acquiring synthetic data known as Continual Class-Specific Impression (CCSI) for previously inaccessible trained classes and presenting a methodology to effectively utilize this data for updating networks when introducing new classes. We obtain CCSI by employing data inversion over gradients of the trained classification model on previous classes starting from the mean image of each class inspired by common landmarks shared among medical images and utilizing continual normalization layers statistics as a regularizer in this pixel-wise optimization process. Subsequently, we update the network by combining the synthesized data with new class data and incorporate several losses, including an intra-domain contrastive loss to generalize the deep network trained on the synthesized data to real data, a margin loss to increase separation among previous classes and new ones, and a cosine-normalized cross-entropy loss to alleviate the adverse effects of imbalanced distributions in training data. Extensive experiments show that the proposed framework achieves state-of-the-art performance on four of the public MedMNIST datasets and in-house echocardiography cine series, with an improvement in classification accuracy of up to 51% compared to baseline data-free methods. Our code is available at https://github.com/ubc-tea/Continual-Impression-CCSI.

AI-based histopathology image analysis reveals a distinct subset of endometrial cancers.

Endometrial cancer (EC) has four molecular subtypes with strong prognostic value and therapeutic implications. The most common subtype (NSMP; No Specific Molecular Profile) is assigned after exclusion of the defining features of the other three molecular subtypes and includes patients with heterogeneous clinical outcomes. In this study, we employ artificial intelligence (AI)-powered histopathology image analysis to differentiate between p53abn and NSMP EC subtypes and consequently identify a sub-group of NSMP EC patients that has markedly inferior progression-free and disease-specific survival (termed 'p53abn-like NSMP'), in a discovery cohort of 368 patients and two independent validation cohorts of 290 and 614 from other centers. Shallow whole genome sequencing reveals a higher burden of copy number abnormalities in the 'p53abn-like NSMP' group compared to NSMP, suggesting that this group is biologically distinct compared to other NSMP ECs. Our work demonstrates the power of AI to detect prognostically different and otherwise unrecognizable subsets of EC where conventional and standard molecular or pathologic criteria fall short, refining image-based tumor classification. This study's findings are applicable exclusively to females.

Toward confident prostate cancer detection using ultrasound: a multi-center study.

PURPOSE: Deep learning-based analysis of micro-ultrasound images to detect cancerous lesions is a promising tool for improving prostate cancer (PCa) diagnosis. An ideal model should confidently identify cancer while responding with appropriate uncertainty when presented with out-of-distribution inputs that arise during deployment due to imaging artifacts and the biological heterogeneity of patients and prostatic tissue. METHODS: Using micro-ultrasound data from 693 patients across 5 clinical centers who underwent micro-ultrasound guided prostate biopsy, we train and evaluate convolutional neural network models for PCa detection. To improve robustness to out-of-distribution inputs, we employ and comprehensively benchmark several state-of-the-art uncertainty estimation methods. RESULTS: PCa detection models achieve performance scores up to 76 % average AUROC with a 10-fold cross validation setup. Models with uncertainty estimation obtain expected calibration error scores as low as 2 % , indicating that confident predictions are very likely to be correct. Visualizations of the model output demonstrate that the model correctly identifies healthy versus malignant tissue. CONCLUSION: Deep learning models have been developed to confidently detect PCa lesions from micro-ultrasound. The performance of these models, determined from a large and diverse dataset, is competitive with visual analysis of magnetic resonance imaging, the clinical benchmark to identify PCa lesions for targeted biopsy. Deep learning with micro-ultrasound should be further studied as an avenue for targeted prostate biopsy.

LensePro: label noise-tolerant prototype-based network for improving cancer detection in prostate ultrasound with limited annotations.

PURPOSE: The standard of care for prostate cancer (PCa) diagnosis is the histopathological analysis of tissue samples obtained via transrectal ultrasound (TRUS) guided biopsy. Models built with deep neural networks (DNNs) hold the potential for direct PCa detection from TRUS, which allows targeted biopsy and subsequently enhances outcomes. Yet, there are ongoing challenges with training robust models, stemming from issues such as noisy labels, out-of-distribution (OOD) data, and limited labeled data. METHODS: This study presents LensePro, a unified method that not only excels in label efficiency but also demonstrates robustness against label noise and OOD data. LensePro comprises two key stages: first, self-supervised learning to extract high-quality feature representations from abundant unlabeled TRUS data and, second, label noise-tolerant prototype-based learning to classify the extracted features. RESULTS: Using data from 124 patients who underwent systematic prostate biopsy, LensePro achieves an AUROC, sensitivity, and specificity of 77.9%, 85.9%, and 57.5%, respectively, for detecting PCa in ultrasound. Our model shows it is effective for detecting OOD data in test time, critical for clinical deployment. Ablation studies demonstrate that each component of our method improves PCa detection by addressing one of the three challenges, reinforcing the benefits of a unified approach. CONCLUSION: Through comprehensive experiments, LensePro demonstrates its state-of-the-art performance for TRUS-based PCa detection. Although further research is necessary to confirm its clinical applicability, LensePro marks a notable advancement in enhancing automated computer-aided systems for detecting prostate cancer in ultrasound.

ImSpect: Image-driven self-supervised learning for surgical margin evaluation with mass spectrometry.

PURPOSE: Real-time assessment of surgical margins is critical for favorable outcomes in cancer patients. The iKnife is a mass spectrometry device that has demonstrated potential for margin detection in cancer surgery. Previous studies have shown that using deep learning on iKnife data can facilitate real-time tissue characterization. However, none of the existing literature on the iKnife facilitate the use of publicly available, state-of-the-art pretrained networks or datasets that have been used in computer vision and other domains. METHODS: In a new framework we call ImSpect, we convert 1D iKnife data, captured during basal cell carcinoma (BCC) surgery, into 2D images in order to capitalize on state-of-the-art image classification networks. We also use self-supervision to leverage large amounts of unlabeled, intraoperative data to accommodate the data requirements of these networks. RESULTS: Through extensive ablation studies, we show that we can surpass previous benchmarks of margin evaluation in BCC surgery using iKnife data, achieving an area under the receiver operating characteristic curve (AUC) of 81%. We also depict the attention maps of the developed DL models to evaluate the biological relevance of the embedding space CONCLUSIONS: We propose a new method for characterizing tissue at the surgical margins, using mass spectrometry data from cancer surgery.

Validation of machine learning models for estimation of left ventricular ejection fraction on point-of-care ultrasound: insights on features that impact performance.

BACKGROUND: Machine learning (ML) algorithms can accurately estimate left ventricular ejection fraction (LVEF) from echocardiography, but their performance on cardiac point-of-care ultrasound (POCUS) is not well understood. OBJECTIVES: We evaluate the performance of an ML model for estimation of LVEF on cardiac POCUS compared with Level III echocardiographers' interpretation and formal echo reported LVEF. METHODS: Clinicians at a tertiary care heart failure clinic prospectively scanned 138 participants using hand-carried devices. Video data were analyzed offline by an ML model for LVEF. We compared the ML model's performance with Level III echocardiographers' interpretation and echo reported LVEF. RESULTS: There were 138 participants scanned, yielding 1257 videos. The ML model generated LVEF predictions on 341 videos. We observed a good intraclass correlation (ICC) between the ML model's predictions and the reference standards (ICC = 0.77-0.84). When comparing LVEF estimates for randomized single POCUS videos, the ICC between the ML model and Level III echocardiographers' estimates was 0.772, and it was 0.778 for videos where quantitative LVEF was feasible. When the Level III echocardiographer reviewed all POCUS videos for a participant, the ICC improved to 0.794 and 0.843 when only accounting for studies that could be segmented. The ML model's LVEF estimates also correlated well with LVEF derived from formal echocardiogram reports (ICC = 0.798). CONCLUSION: Our results suggest that clinician-driven cardiac POCUS produces ML model LVEF estimates that correlate well with expert interpretation and echo reported LVEF.

Automated Atrial Fibrillation Diagnosis by Echocardiography without ECG: Accuracy and Applications of a New Deep Learning Approach.

BACKGROUND: Automated rhythm detection on echocardiography through artificial intelligence (AI) has yet to be fully realized. We propose an AI model trained to identify atrial fibrillation (AF) using apical 4-chamber (AP4) cines without requiring electrocardiogram (ECG) data. METHODS: Transthoracic echocardiography studies of consecutive patients ≥ 18 years old at our tertiary care centre were retrospectively reviewed for AF and sinus rhythm. The study was first interpreted by level III-trained echocardiography cardiologists as the gold standard for rhythm diagnosis based on ECG rhythm strip and imaging assessment, which was also verified with a 12-lead ECG around the time of the study. AP4 cines with three cardiac cycles were then extracted from these studies with the rhythm strip and Doppler information removed and introduced to the deep learning model ResNet(2+1)D with an 80:10:10 training-validation-test split ratio. RESULTS: 634 patient studies (1205 cines) were included. After training, the AI model achieved high accuracy on validation for detection of both AF and sinus rhythm (mean F1-score = 0.92; AUROC = 0.95). Performance was consistent on the test dataset (mean F1-score = 0.94, AUROC = 0.98) when using the cardiologist's assessment of the ECG rhythm strip as the gold standard, who had access to the full study and external ECG data, while the AI model did not. CONCLUSIONS: AF detection by AI on echocardiography without ECG appears accurate when compared to an echocardiography cardiologist's assessment of the ECG rhythm strip as the gold standard. This has potential clinical implications in point-of-care ultrasound and stroke risk stratification.

Self-Supervised Learning With Limited Labeled Data for Prostate Cancer Detection in High-Frequency Ultrasound.

Deep learning-based analysis of high-frequency, high-resolution micro-ultrasound data shows great promise for prostate cancer (PCa) detection. Previous approaches to analysis of ultrasound data largely follow a supervised learning (SL) paradigm. Ground truth labels for ultrasound images used for training deep networks often include coarse annotations generated from the histopathological analysis of tissue samples obtained via biopsy. This creates inherent limitations on the availability and quality of labeled data, posing major challenges to the success of SL methods. However, unlabeled prostate ultrasound data are more abundant. In this work, we successfully apply self-supervised representation learning to micro-ultrasound data. Using ultrasound data from 1028 biopsy cores of 391 subjects obtained in two clinical centers, we demonstrate that feature representations learned with this method can be used to classify cancer from noncancer tissue, obtaining an AUROC score of 91% on an independent test set. To the best of our knowledge, this is the first successful end-to-end self-SL (SSL) approach for PCa detection using ultrasound data. Our method outperforms baseline SL approaches, generalizes well between different data centers, and scales well in performance as more unlabeled data are added, making it a promising approach for future research using large volumes of unlabeled data. Our code is publicly available at https://www.github.com/MahdiGilany/SSL_micro_ultrasound.

Cardiac Phenotyping of SARS-CoV-2 in British Columbia: A Prospective Echo Study With Strain Imaging.

BACKGROUND: There is limited data on the residual echocardiographic findings including strain analysis among post-coronavirus disease (COVID) patients. The aim of our study is to prospectively phenotype post-COVID patients. METHODS: All patients discharged following acute COVID infection were systematically followed in the post-COVID-19 Recovery Clinic at Vancouver General Hospital and St. Paul's Hospital. At 4-18 weeks post diagnosis, patients underwent comprehensive echocardiographic assessment. Left ventricular ejection fraction (LVEF) was assessed by 3D, 2D Biplane Simpson's, or visual estimate. LV global longitudinal strain (GLS) was measured using a vendor-independent 2D speckle-tracking software (TomTec). RESULTS: A total of 127 patients (53% female, mean age 58 years) were included in our analyses. At baseline, cardiac conditions were present in 58% of the patients (15% coronary artery disease, 4% heart failure, 44% hypertension, 10% atrial fibrillation) while the remainder were free of cardiac conditions. COVID-19 serious complications were present in 79% of the patients (76% pneumonia, 37% intensive care unit admission, 21% intubation, 1% myocarditis). Normal LVEF was seen in 96% of the cohort and 97% had normal right ventricular systolic function. A high proportion (53%) had abnormal LV GLS defined as < 18%. Average LV GLS of septal and inferior segments were lower compared to that of other segments. Among patients without pre-existing cardiac conditions, LVEF was abnormal in only 1.9%, but LV GLS was abnormal in 46% of the patients. CONCLUSIONS: Most post-COVID patients had normal LVEF at 4-18 weeks post diagnosis, but over half had abnormal LV GLS.

Deep Learning for Detection and Localization of B-Lines in Lung Ultrasound.

Lung ultrasound (LUS) is an important imaging modality used by emergency physicians to assess pulmonary congestion at the patient bedside. B-line artifacts in LUS videos are key findings associated with pulmonary congestion. Not only can the interpretation of LUS be challenging for novice operators, but visual quantification of B-lines remains subject to observer variability. In this work, we investigate the strengths and weaknesses of multiple deep learning approaches for automated B-line detection and localization in LUS videos. We curate and publish, BEDLUS, a new ultrasound dataset comprising 1,419 videos from 113 patients with a total of 15,755 expert-annotated B-lines. Based on this dataset, we present a benchmark of established deep learning methods applied to the task of B-line detection. To pave the way for interpretable quantification of B-lines, we propose a novel "single-point" approach to B-line localization using only the point of origin. Our results show that (a) the area under the receiver operating characteristic curve ranges from 0.864 to 0.955 for the benchmarked detection methods, (b) within this range, the best performance is achieved by models that leverage multiple successive frames as input, and (c) the proposed single-point approach for B-line localization reaches an F 1-score of 0.65, performing on par with the inter-observer agreement. The dataset and developed methods can facilitate further biomedical research on automated interpretation of lung ultrasound with the potential to expand the clinical utility.

TRUSformer: improving prostate cancer detection from micro-ultrasound using attention and self-supervision.

PURPOSE: A large body of previous machine learning methods for ultrasound-based prostate cancer detection classify small regions of interest (ROIs) of ultrasound signals that lie within a larger needle trace corresponding to a prostate tissue biopsy (called biopsy core). These ROI-scale models suffer from weak labeling as histopathology results available for biopsy cores only approximate the distribution of cancer in the ROIs. ROI-scale models do not take advantage of contextual information that are normally considered by pathologists, i.e., they do not consider information about surrounding tissue and larger-scale trends when identifying cancer. We aim to improve cancer detection by taking a multi-scale, i.e., ROI-scale and biopsy core-scale, approach. METHODS: Our multi-scale approach combines (i) an "ROI-scale" model trained using self-supervised learning to extract features from small ROIs and (ii) a "core-scale" transformer model that processes a collection of extracted features from multiple ROIs in the needle trace region to predict the tissue type of the corresponding core. Attention maps, as a by-product, allow us to localize cancer at the ROI scale. RESULTS: We analyze this method using a dataset of micro-ultrasound acquired from 578 patients who underwent prostate biopsy, and compare our model to baseline models and other large-scale studies in the literature. Our model shows consistent and substantial performance improvements compared to ROI-scale-only models. It achieves [Formula: see text] AUROC, a statistically significant improvement over ROI-scale classification. We also compare our method to large studies on prostate cancer detection, using other imaging modalities. CONCLUSIONS: Taking a multi-scale approach that leverages contextual information improves prostate cancer detection compared to ROI-scale-only models. The proposed model achieves a statistically significant improvement in performance and outperforms other large-scale studies in the literature. Our code is publicly available at www.github.com/med-i-lab/TRUSFormer .

Machine learning derived echocardiographic image quality in patients with left ventricular systolic dysfunction: insights on the echo views of greatest image quality.

We sought to determine the cardiac ultrasound view of greatest quality using a machine learning (ML) approach on a cohort of transthoracic echocardiograms (TTE) with abnormal left ventricular (LV) systolic function. We utilize an ML model to determine the TTE view of highest quality when scanned by sonographers. A random sample of TTEs with reported LV dysfunction from 09/25/2017-01/15/2019 were downloaded from the regional database. Component video files were analyzed using ML models that jointly classified view and image quality. The model consisted of convolutional layers for extracting spatial features and Long Short-term Memory units to temporally aggregate the frame-wise spatial embeddings. We report the view-specific quality scores for each TTE. Pair-wise comparisons amongst views were performed with Wilcoxon signed-rank test. Of 1,145 TTEs analyzed by the ML model, 74.5% were from males and mean LV ejection fraction was 43.1 ± 9.9%. Maximum quality score was best for the apical 4 chamber (AP4) view (70.6 ± 13.9%, p<0.001 compared to all other views) and worst for the apical 2 chamber (AP2) view (60.4 ± 15.4%, p<0.001 for all views except parasternal short-axis view at mitral/papillary muscle level, PSAX M/PM). In TTEs scanned by professional sonographers, the view with greatest ML-derived quality was the AP4 view.

Quantifying deep neural network uncertainty for atrial fibrillation detection with limited labels.

Atrial fibrillation (AF) is the most common arrhythmia found in the intensive care unit (ICU), and is associated with many adverse outcomes. Effective handling of AF and similar arrhythmias is a vital part of modern critical care, but obtaining knowledge about both disease burden and effective interventions often requires costly clinical trials. A wealth of continuous, high frequency physiological data such as the waveforms derived from electrocardiogram telemetry are promising sources for enriching clinical research. Automated detection using machine learning and in particular deep learning has been explored as a solution for processing these data. However, a lack of labels, increased presence of noise, and inability to assess the quality and trustworthiness of many machine learning model predictions pose challenges to interpretation. In this work, we propose an approach for training deep AF models on limited, noisy data and report uncertainty in their predictions. Using techniques from the fields of weakly supervised learning, we leverage a surrogate model trained on non-ICU data to create imperfect labels for a large ICU telemetry dataset. We combine these weak labels with techniques to estimate model uncertainty without the need for extensive human data annotation. AF detection models trained using this process demonstrated higher classification performance (0.64-0.67 F1 score) and improved calibration (0.05-0.07 expected calibration error).

Uncertainty estimation for margin detection in cancer surgery using mass spectrometry.

PURPOSE: Rapid evaporative ionization mass spectrometry (REIMS) is an emerging technology for clinical margin detection. Deployment of REIMS depends on construction of reliable deep learning models that can categorize tissue according to its metabolomic signature. Challenges associated with developing these models include the presence of noise during data acquisition and the variance in tissue signatures between patients. In this study, we propose integration of uncertainty estimation in deep models to factor predictive confidence into margin detection in cancer surgery. METHODS: iKnife is used to collect 693 spectra of cancer and healthy samples acquired from 91 patients during basal cell carcinoma resection. A Bayesian neural network and two baseline models are trained on these data to perform classification as well as uncertainty estimation. The samples with high estimated uncertainty are then removed, and new models are trained using the clean data. The performance of proposed and baseline models, with different ratios of filtered data, is then compared. RESULTS: The data filtering does not improve the performance of the baseline models as they cannot provide reliable estimations of uncertainty. In comparison, the proposed model demonstrates a statistically significant improvement in average balanced accuracy (75.2%), sensitivity (74.1%) and AUC (82.1%) after removing uncertain training samples. We also demonstrate that if highly uncertain samples are predicted and removed from the test data, sensitivity further improves to 88.2%. CONCLUSIONS: This is the first study that applies uncertainty estimation to inform model training and deployment for tissue recognition in cancer surgery. Uncertainty estimation is leveraged in two ways: by factoring a measure of input noise in training the models and by including predictive confidence in reporting the outputs. We empirically show that considering uncertainty for model development can help improve the overall accuracy of a margin detection system using REIMS.

A novel ultrasound software system for lumbar level identification in obstetric patients.

PURPOSE: Using machine learning, we developed a proprietary ultrasound software called the Spine Level Identification (SLIDE) system, which automatically identifies lumbar landmarks in real time as the operator slides the transducer over the lumber spine. Here, we assessed the agreement between SLIDE and manual palpation and traditional lumbar ultrasound (LUS) for determining the primary target L3-4 interspace. METHODS: Upon institutional ethics approval and informed consent, 76 healthy term parturients scheduled for elective Caesarean delivery were recruited. The L3-4 interspace was identified by manual palpation and then by the SLIDE method. The reference standard was located using traditional LUS by an experienced operator. The primary outcome was the L3-4 interspace identification agreement of manual palpation and SLIDE with the reference standard, as percentage agreement and Gwet's agreement coefficient (AC1). RESULTS: The raw agreement was 70% with Gwet's agreement coefficient (AC1) = 0.59 (95% confidence interval [CI], 0.41 to 0.77) for manual palpation and 84% with Gwet's AC1 = 0.82 (95% CI, 0.70 to 0.93) for SLIDE. When the levels differ from the reference, the manual palpation method identified L2-3 more often than L4-5 while the SLIDE method identified equally above or below L3-4. The SLIDE system had greater agreement than palpation in locating L3-4 and all other lumber interspaces after controlling for body mass index (adjusted odds ratio, 2.99; 95% CI, 1.21 to 8.7; P = 0.02). CONCLUSION: The SLIDE system had higher agreement with traditional ultrasound than manual palpation did in identifying L3-4 and all other lumber interspaces after adjusting for BMI in healthy term obstetric patients. Future studies should examine factors that affect agreement and ways to improve SLIDE for clinical integration. STUDY REGISTRATION: www. CLINICALTRIALS: gov (NCT02982317); registered 5 December 2016.

RéSUMé: OBJECTIF: À l'aide de l'apprentissage automatique, nous avons développé un logiciel d'échographie propriétaire appelé SLIDE (pour Spine Level Identification, c.-à-d. système d'identification du niveau vertébral), qui identifie automatiquement les points de repère lombaires en temps réel lorsque l'opérateur fait passer le transducteur sur la colonne lombaire. Ici, nous avons évalué l'agrément entre le SLIDE et la palpation manuelle et l'échographie lombaire traditionnelle pour déterminer l'espace intervertébral cible principal L3­L4. MéTHODE: Après avoir obtenu l'approbation du comité d'éthique de l'établissement et le consentement éclairé, 76 parturientes en bonne santé et à terme devant bénéficier d'un accouchement par césarienne programmée ont été recrutées. L'espace intervertébral L3­L4 a été identifié par palpation manuelle puis avec le logiciel SLIDE. L'étalon de référence a été localisé à l'aide d'une échographie lombaire traditionnelle par un opérateur expérimenté. Le critère d'évaluation principal était l'agrément entre l'identification de l'espace intervertébral L3­L4 par palpation manuelle et par logiciel SLIDE avec l'étalon de référence, en pourcentage d'agrément et coefficient d'agrément de Gwet (CA1). RéSULTATS: L'agrément brut était de 70 % avec le coefficient d'agrément de Gwet (CA1) = 0,59 (intervalle de confiance [IC] à 95 %, 0,41 à 0,77) pour la palpation manuelle et de 84 % avec le CA1 de Gwet = 0,82 (IC 95 %, 0,70 à 0,93) pour le logiciel SLIDE. Lorsque les niveaux lombaires différaient de la référence, la méthode de palpation manuelle a identifié L2­L3 plus souvent que L4­L5, tandis que la méthode SLIDE a identifié les vertèbres supérieures ou inférieures à L3­L4 de manière égale. Le système SLIDE a affiché un agrément plus important que la palpation pour localiser L3­L4 et tous les autres espaces intervertébraux lombaires après ajustement pour tenir compte de l'indice de masse corporelle (rapport de cotes ajusté, 2,99; IC 95 %, 1,21 à 8,7; P = 0,02). CONCLUSION: Le système SLIDE avait affiché un agrément plus élevé avec l'échographie traditionnelle que la palpation manuelle pour identifier le niveau L3­L4 et tous les autres espaces intervertébraux lombaires après ajustement pour tenir compte de l'IMC chez les patientes obstétricales à terme en bonne santé. Une étude future devrait examiner les facteurs qui affectent l'agrément et les moyens d'améliorer le logiciel SLIDE pour une intégration clinique. ENREGISTREMENT DE L'éTUDE: www.clinicaltrials.gov (NCT02982317); enregistrée le 5 décembre 2016.

Training deep neural networks with noisy clinical labels: toward accurate detection of prostate cancer in US data.

PURPOSE: Ultrasound is the standard-of-care to guide the systematic biopsy of the prostate. During the biopsy procedure, up to 12 biopsy cores are randomly sampled from six zones within the prostate, where the histopathology of those cores is used to determine the presence and grade of the cancer. Histopathology reports only provide statistical information on the presence of cancer and do not normally contain fine-grain information of cancer distribution within each core. This limitation hinders the development of machine learning models to detect the presence of cancer in ultrasound so that biopsy can be more targeted to highly suspicious prostate regions. METHODS: In this paper, we tackle this challenge in the form of training with noisy labels derived from histopathology. Noisy labels often result in the model overfitting to the training data, hence limiting its generalizability. To avoid overfitting, we focus on the generalization of the features of the model and present an iterative data label refinement algorithm to amend the labels gradually. We simultaneously train two classifiers, with the same structure, and automatically stop the training when we observe any sign of overfitting. Then, we use a confident learning approach to clean the data labels and continue with the training. This process is iteratively applied to the training data and labels until convergence. RESULTS: We illustrate the performance of the proposed method by classifying prostate cancer using a dataset of ultrasound images from 353 biopsy cores obtained from 90 patients. We achieve area under the curve, sensitivity, specificity, and accuracy of 0.73, 0.80, 0.63, and 0.69, respectively. CONCLUSION: Our approach is able to provide clinicians with a visualization of regions that likely contain cancerous tissue to obtain more accurate biopsy samples. The results demonstrate that our proposed method produces superior accuracy compared to the state-of-the-art methods.

A Novel Continuous Left Ventricular Diastolic Function Score Using Machine Learning.

BACKGROUND: Unlike left ventricular (LV) ejection fraction, which provides a precise, reliable, and prognostically valuable measure of systolic function, there is no single analogous measure of LV diastolic function. OBJECTIVES: We aimed to develop a continuous score to grade LV diastolic function using machine learning modeling of echocardiographic data. METHODS: Consecutive echo studies performed at a tertiary-care center between February 1, 2010, and March 31, 2016, were assessed, excluding studies containing features that would interfere with diastolic function assessment as well as studies in which 1 or more parameters within the contemporary diastolic function assessment algorithm were not reported. Diastolic function was graded based on 2016 American Society of Echocardiography (ASE)/European Association of Cardiovascular Imaging (EACVI) guidelines, excluding indeterminate studies. Machine learning models were trained (support vector machine [SVM], decision tree [DT], XGBoost [XGB], and dense neural network [DNN]) to classify studies within the training set by diastolic dysfunction severity, blinded to the ASE/EACVI classification. The DNN model was retrained to generate a regression model (R-DNN) to predict a continuous LV diastolic function score. RESULTS: A total of 28,986 studies were included; 23,188 studies were used to train the models, and 5,798 studies were used for validation. The models were able to reclassify studies with high agreement to the ASE/EACVI algorithm (SVM, 83%; DT, 100%; XGB, 100%; DNN, 98%). The continuous diastolic function score corresponded well with ASE/EACVI guidelines, with scores of 1.00 ± 0.01 for studies with normal function and 0.74 ± 0.05, 0.51 ± 0.06, and 0.27 ± 0.11 for mild, moderate, and severe diastolic dysfunction, respectively (mean ± 1 SD). A score of <0.91 predicted abnormal diastolic function (area under the receiver operator curve = 0.99), while a score of <0.65 predicted elevated filling pressure (area under the receiver operator curve = 0.99). CONCLUSIONS: Machine learning can assimilate echocardiographic data and generate an automated continuous diastolic function score that corresponds well with current diastolic function grading recommendations.

Coarse label refinement for improving prostate cancer detection in ultrasound imaging.

PURPOSE: Ultrasound-guided biopsy plays a major role in prostate cancer (PCa) detection, yet is limited by a high rate of false negatives and low diagnostic yield of the current systematic, non-targeted approaches. Developing machine learning models for accurately identifying cancerous tissue in ultrasound would help sample tissues from regions with higher cancer likelihood. A plausible approach for this purpose is to use individual ultrasound signals corresponding to a core as inputs and consider the histopathology diagnosis for the entire core as labels. However, this introduces significant amount of label noise to training and degrades the classification performance. Previously, we suggested that histopathology-reported cancer involvement can be a reasonable approximation for the label noise. METHODS: Here, we propose an involvement-based label refinement (iLR) method to correct corrupted labels and improve cancer classification. The difference between predicted and true cancer involvements is used to guide the label refinement process. We further incorporate iLR into state-of-the-art methods for learning with noisy labels and predicting cancer involvement. RESULTS: We use 258 biopsy cores from 70 patients and demonstrate that our proposed label refinement method improves the performance of multiple noise-tolerant approaches and achieves a balanced accuracy, correlation coefficient, and mean absolute error of 76.7%, 0.68, and 12.4, respectively. CONCLUSIONS: Our key contribution is to leverage a data-centric method to deal with noisy labels using histopathology reports, and improve the performance of prostate cancer diagnosis through a hierarchical training process with label refinement.

U-LanD: Uncertainty-Driven Video Landmark Detection.

This paper presents U-LanD, a framework for automatic detection of landmarks on key frames of the video by leveraging the uncertainty of landmark prediction. We tackle a specifically challenging problem, where training labels are noisy and highly sparse. U-LanD builds upon a pivotal observation: a deep Bayesian landmark detector solely trained on key video frames, has significantly lower predictive uncertainty on those frames vs. other frames in videos. We use this observation as an unsupervised signal to automatically recognize key frames on which we detect landmarks. As a test-bed for our framework, we use ultrasound imaging videos of the heart, where sparse and noisy clinical labels are only available for a single frame in each video. Using data from 4,493 patients, we demonstrate that U-LanD can exceedingly outperform the state-of-the-art non-Bayesian counterpart by a noticeable absolute margin of 42% in R2 score, with almost no overhead imposed on the model size.