Tissue Contamination Challenges the Credibility of Machine Learning Models in Real World Digital Pathology.

Machine learning (ML) models are poised to transform surgical pathology practice. The most successful use attention mechanisms to examine whole slides, identify which areas of tissue are diagnostic, and use them to guide diagnosis. Tissue contaminants, such as floaters, represent unexpected tissue. Although human pathologists are extensively trained to consider and detect tissue contaminants, we examined their impact on ML models. We trained 4 whole-slide models. Three operate in placenta for the following functions: (1) detection of decidual arteriopathy, (2) estimation of gestational age, and (3) classification of macroscopic placental lesions. We also developed a model to detect prostate cancer in needle biopsies. We designed experiments wherein patches of contaminant tissue are randomly sampled from known slides and digitally added to patient slides and measured model performance. We measured the proportion of attention given to contaminants and examined the impact of contaminants in the t-distributed stochastic neighbor embedding feature space. Every model showed performance degradation in response to one or more tissue contaminants. Decidual arteriopathy detection--balanced accuracy decreased from 0.74 to 0.69 ± 0.01 with addition of 1 patch of prostate tissue for every 100 patches of placenta (1% contaminant). Bladder, added at 10% contaminant, raised the mean absolute error in estimating gestational age from 1.626 weeks to 2.371 ± 0.003 weeks. Blood, incorporated into placental sections, induced false-negative diagnoses of intervillous thrombi. Addition of bladder to prostate cancer needle biopsies induced false positives, a selection of high-attention patches, representing 0.033 mm2, and resulted in a 97% false-positive rate when added to needle biopsies. Contaminant patches received attention at or above the rate of the average patch of patient tissue. Tissue contaminants induce errors in modern ML models. The high level of attention given to contaminants indicates a failure to encode biological phenomena. Practitioners should move to quantify and ameliorate this problem.

Tissue contamination challenges the credibility of machine learning models in real world digital pathology.

Machine learning (ML) models are poised to transform surgical pathology practice. The most successful use attention mechanisms to examine whole slides, identify which areas of tissue are diagnostic, and use them to guide diagnosis. Tissue contaminants, such as floaters, represent unexpected tissue. While human pathologists are extensively trained to consider and detect tissue contaminants, we examined their impact on ML models. We trained 4 whole slide models. Three operate in placenta for 1) detection of decidual arteriopathy (DA), 2) estimation of gestational age (GA), and 3) classification of macroscopic placental lesions. We also developed a model to detect prostate cancer in needle biopsies. We designed experiments wherein patches of contaminant tissue are randomly sampled from known slides and digitally added to patient slides and measured model performance. We measured the proportion of attention given to contaminants and examined the impact of contaminants in T-distributed Stochastic Neighbor Embedding (tSNE) feature space. Every model showed performance degradation in response to one or more tissue contaminants. DA detection balanced accuracy decreased from 0.74 to 0.69 +/- 0.01 with addition of 1 patch of prostate tissue for every 100 patches of placenta (1% contaminant). Bladder, added at 10% contaminant raised the mean absolute error in estimating gestation age from 1.626 weeks to 2.371 +/ 0.003 weeks. Blood, incorporated into placental sections, induced false negative diagnoses of intervillous thrombi. Addition of bladder to prostate cancer needle biopsies induced false positives, a selection of high-attention patches, representing 0.033mm2, resulted in a 97% false positive rate when added to needle biopsies. Contaminant patches received attention at or above the rate of the average patch of patient tissue. Tissue contaminants induce errors in modern ML models. The high level of attention given to contaminants indicates a failure to encode biological phenomena. Practitioners should move to quantify and ameliorate this problem.

Machine learning classification of placental villous infarction, perivillous fibrin deposition, and intervillous thrombus.

INTRODUCTION: Placental parenchymal lesions are commonly encountered and carry significant clinical associations. However, they are frequently missed or misclassified by general practice pathologists. Interpretation of pathology slides has emerged as one of the most successful applications of machine learning (ML) in medicine with applications ranging from cancer detection and prognostication to transplant medicine. The goal of this study was to use a whole-slide learning model to identify and classify placental parenchymal lesions including villous infarctions, intervillous thrombi (IVT), and perivillous fibrin deposition (PVFD). METHODS: We generated whole slide images from placental discs examined at our institution with infarct, IVT, PVFD, or no macroscopic lesion. Slides were analyzed as a set of overlapping patches. We extracted feature vectors from each patch using a pretrained convolutional neural network (EfficientNetV2L). We trained a model to assign attention to each vector and used the attentions as weights to produce a pooled feature vector. The pooled vector was classified as normal or 1 of 3 lesions using a fully connected network. Patch attention was plotted to highlight informative areas of the slide. RESULTS: Overall balanced accuracy in a test set of held-out slides was 0.86 with receiver-operator characteristic areas under the curve of 0.917-0.993. Cases of PVFD were frequently miscalled as normal or infarcts, the latter possibly due to the perivillous fibrin found at the periphery of infarctions. We used attention maps to further understand some errors, including one most likely due to poor tissue fixation and processing. DISCUSSION: We used a whole-slide learning paradigm to train models to recognize three of the most common placental parenchymal lesions. We used attention maps to gain insight into model function, which differed from intuitive explanations.

Multi-Contrast MRI Segmentation Trained on Synthetic Images.

In our comprehensive experiments and evaluations, we show that it is possible to generate multiple contrast (even all synthetically) and use synthetically generated images to train an image segmentation engine. We showed promising segmentation results tested on real multi-contrast MRI scans when delineating muscle, fat, bone and bone marrow, all trained on synthetic images. Based on synthetic image training, our segmentation results were as high as 93.91%, 94.11%, 91.63%, 95.33%, for muscle, fat, bone, and bone marrow delineation, respectively. Results were not significantly different from the ones obtained when real images were used for segmentation training: 94.68%, 94.67%, 95.91%, and 96.82%, respectively. Clinical relevance- Synthetically generated images could potentially be used in large-scale training of deep networks for segmentation purpose. Small data set problem of many clinical imaging problems can potentially be addressed with the proposed algorithm.

Information Bottleneck Attribution for Visual Explanations of Diagnosis and Prognosis.

Visual explanation methods have an important role in the prognosis of the patients where the annotated data is limited or unavailable. There have been several attempts to use gradient-based attribution methods to localize pathology from medical scans without using segmentation labels. This research direction has been impeded by the lack of robustness and reliability. These methods are highly sensitive to the network parameters. In this study, we introduce a robust visual explanation method to address this problem for medical applications. We provide an innovative visual explanation algorithm for general purpose and as an example application we demonstrate its effectiveness for quantifying lesions in the lungs caused by the Covid-19 with high accuracy and robustness without using dense segmentation labels. This approach overcomes the drawbacks of commonly used Grad-CAM and its extended versions. The premise behind our proposed strategy is that the information flow is minimized while ensuring the classifier prediction stays similar. Our findings indicate that the bottleneck condition provides a more stable severity estimation than the similar attribution methods. The source code will be publicly available upon publication.

Hierarchical 3D Feature Learning for Pancreas Segmentation.

We propose a novel 3D fully convolutional deep network for automated pancreas segmentation from both MRI and CT scans. More specifically, the proposed model consists of a 3D encoder that learns to extract volume features at different scales; features taken at different points of the encoder hierarchy are then sent to multiple 3D decoders that individually predict intermediate segmentation maps. Finally, all segmentation maps are combined to obtain a unique detailed segmentation mask. We test our model on both CT and MRI imaging data: the publicly available NIH Pancreas-CT dataset (consisting of 82 contrast-enhanced CTs) and a private MRI dataset (consisting of 40 MRI scans). Experimental results show that our model outperforms existing methods on CT pancreas segmentation, obtaining an average Dice score of about 88%, and yields promising segmentation performance on a very challenging MRI data set (average Dice score is about 77%). Additional control experiments demonstrate that the achieved performance is due to the combination of our 3D fully-convolutional deep network and the hierarchical representation decoding, thus substantiating our architectural design.

Capsules for biomedical image segmentation.

Our work expands the use of capsule networks to the task of object segmentation for the first time in the literature. This is made possible via the introduction of locally-constrained routing and transformation matrix sharing, which reduces the parameter/memory burden and allows for the segmentation of objects at large resolutions. To compensate for the loss of global information in constraining the routing, we propose the concept of "deconvolutional" capsules to create a deep encoder-decoder style network, called SegCaps. We extend the masked reconstruction regularization to the task of segmentation and perform thorough ablation experiments on each component of our method. The proposed convolutional-deconvolutional capsule network, SegCaps, shows state-of-the-art results while using a fraction of the parameters of popular segmentation networks. To validate our proposed method, we perform experiments segmenting pathological lungs from clinical and pre-clinical thoracic computed tomography (CT) scans and segmenting muscle and adipose (fat) tissue from magnetic resonance imaging (MRI) scans of human subjects' thighs. Notably, our experiments in lung segmentation represent the largest-scale study in pathological lung segmentation in the literature, where we conduct experiments across five extremely challenging datasets, containing both clinical and pre-clinical subjects, and nearly 2000 computed-tomography scans. Our newly developed segmentation platform outperforms other methods across all datasets while utilizing less than 5% of the parameters in the popular U-Net for biomedical image segmentation. Further, we demonstrate capsules' ability to generalize to unseen handling of rotations/reflections on natural images.

A Novel Extension to Fuzzy Connectivity for Body Composition Analysis: Applications in Thigh, Brain, and Whole Body Tissue Segmentation.

Magnetic resonance imaging (MRI) is the non-invasive modality of choice for body tissue composition analysis due to its excellent soft-tissue contrast and lack of ionizing radiation. However, quantification of body composition requires an accurate segmentation of fat, muscle, and other tissues from MR images, which remains a challenging goal due to the intensity overlap between them. In this study, we propose a fully automated, data-driven image segmentation platform that addresses multiple difficulties in segmenting MR images such as varying inhomogeneity, non-standardness, and noise, while producing a high-quality definition of different tissues. In contrast to most approaches in the literature, we perform segmentation operation by combining three different MRI contrasts and a novel segmentation tool, which takes into account variability in the data. The proposed system, based on a novel affinity definition within the fuzzy connectivity image segmentation family, prevents the need for user intervention and reparametrization of the segmentation algorithms. In order to make the whole system fully automated, we adapt an affinity propagation clustering algorithm to roughly identify tissue regions and image background. We perform a thorough evaluation of the proposed algorithm's individual steps as well as comparison with several approaches from the literature for the main application of muscle/fat separation. Furthermore, whole-body tissue composition and brain tissue delineation were conducted to show the generalization ability of the proposed system. This new automated platform outperforms other state-of-the-art segmentation approaches both in accuracy and efficiency.