Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer.

Importance: Application of deep learning algorithms to whole-slide pathology images can potentially improve diagnostic accuracy and efficiency. Objective: Assess the performance of automated deep learning algorithms at detecting metastases in hematoxylin and eosin-stained tissue sections of lymph nodes of women with breast cancer and compare it with pathologists' diagnoses in a diagnostic setting. Design, Setting, and Participants: Researcher challenge competition (CAMELYON16) to develop automated solutions for detecting lymph node metastases (November 2015-November 2016). A training data set of whole-slide images from 2 centers in the Netherlands with (n = 110) and without (n = 160) nodal metastases verified by immunohistochemical staining were provided to challenge participants to build algorithms. Algorithm performance was evaluated in an independent test set of 129 whole-slide images (49 with and 80 without metastases). The same test set of corresponding glass slides was also evaluated by a panel of 11 pathologists with time constraint (WTC) from the Netherlands to ascertain likelihood of nodal metastases for each slide in a flexible 2-hour session, simulating routine pathology workflow, and by 1 pathologist without time constraint (WOTC). Exposures: Deep learning algorithms submitted as part of a challenge competition or pathologist interpretation. Main Outcomes and Measures: The presence of specific metastatic foci and the absence vs presence of lymph node metastasis in a slide or image using receiver operating characteristic curve analysis. The 11 pathologists participating in the simulation exercise rated their diagnostic confidence as definitely normal, probably normal, equivocal, probably tumor, or definitely tumor. Results: The area under the receiver operating characteristic curve (AUC) for the algorithms ranged from 0.556 to 0.994. The top-performing algorithm achieved a lesion-level, true-positive fraction comparable with that of the pathologist WOTC (72.4% [95% CI, 64.3%-80.4%]) at a mean of 0.0125 false-positives per normal whole-slide image. For the whole-slide image classification task, the best algorithm (AUC, 0.994 [95% CI, 0.983-0.999]) performed significantly better than the pathologists WTC in a diagnostic simulation (mean AUC, 0.810 [range, 0.738-0.884]; P < .001). The top 5 algorithms had a mean AUC that was comparable with the pathologist interpreting the slides in the absence of time constraints (mean AUC, 0.960 [range, 0.923-0.994] for the top 5 algorithms vs 0.966 [95% CI, 0.927-0.998] for the pathologist WOTC). Conclusions and Relevance: In the setting of a challenge competition, some deep learning algorithms achieved better diagnostic performance than a panel of 11 pathologists participating in a simulation exercise designed to mimic routine pathology workflow; algorithm performance was comparable with an expert pathologist interpreting whole-slide images without time constraints. Whether this approach has clinical utility will require evaluation in a clinical setting.

Ensemble of convolutional neural networks for classification of breast microcalcification from mammograms.

Human level recall performance in detecting breast cancer considering microcalcifications from mammograms has a recall value between 74.5% and 92.3%. In this research, we approach to breast microcalcification classification problem using convolutional neural networks along with various preprocessing methods such as contrast scaling, dilation, cropping etc. and decision fusion using ensemble of networks. Various experiments on Digital Database for Screening Mammography dataset showed that preprocessing poses great importance on the classification performance. The stand-alone models using the dilation and cropping preprocessing techniques achieved the highest recall value of 91.3%. The ensembles of the stand-alone models surpass this recall value and a 97.3% value of recall is achieved. The ensemble having the highest F1 Score (harmonic mean of precision and recall), which is 94.5%, has a recall value of 94.0% and a precision value of 95.0%. This recall is still above human level performance and the models achieve competitive results in terms of accuracy, precision, recall and F1 score measures.

A novel deep learning approach for classification of EEG motor imagery signals.

OBJECTIVE: Signal classification is an important issue in brain computer interface (BCI) systems. Deep learning approaches have been used successfully in many recent studies to learn features and classify different types of data. However, the number of studies that employ these approaches on BCI applications is very limited. In this study we aim to use deep learning methods to improve classification performance of EEG motor imagery signals. APPROACH: In this study we investigate convolutional neural networks (CNN) and stacked autoencoders (SAE) to classify EEG Motor Imagery signals. A new form of input is introduced to combine time, frequency and location information extracted from EEG signal and it is used in CNN having one 1D convolutional and one max-pooling layers. We also proposed a new deep network by combining CNN and SAE. In this network, the features that are extracted in CNN are classified through the deep network SAE. MAIN RESULTS: The classification performance obtained by the proposed method on BCI competition IV dataset 2b in terms of kappa value is 0.547. Our approach yields 9% improvement over the winner algorithm of the competition. SIGNIFICANCE: Our results show that deep learning methods provide better classification performance compared to other state of art approaches. These methods can be applied successfully to BCI systems where the amount of data is large due to daily recording.

Applying Computational Aesthetics to a Video Game Application Using Machine Learning.

The authors have developed a novel approach to evaluating the aesthetic quality of the camera direction in video game scenes rendered in real time while the game is being played. Their goal was to improve the visual aesthetic quality of computer-generated images using a computational aesthetics approach via a regression machine learning model. Considering the challenges and limitations involved, the proposed approach yielded promising prediction performance. The results show that near-real-time aesthetic analysis and visual improvement is possible using a virtual camera director.

A comparative analysis of functional connectivity data in resting and task-related conditions of the brain for disease signature of OCD.

Obsessive Compulsive Disorder (OCD) is a frequent, chronic disorder producing intrusive thoughts which results in repetitive behaviors. It is thought that this psychological disorder occurs due to abnormal functional connectivity in certain regions of the brain called Default Mode Network (DMN) mainly. Recently, functional MRI (FMRI) studies were performed in order to compare the differences in brain activity between patients with OCD and healthy individuals through different conditions of the brain. Our previous study on extraction of disease signature for OCD that is determining the features for discrimination of OCD patients from healthy individuals based on their resting-sate functional connectivity (rs-FC) data had given encouraging results. In the present study, functional data extracted from FMRI images of subjects under imagination task (maintaining an image in mind, im-FC) is considered. The aim of this study is to compare classification results achieved from both resting and task-related (imagination) conditions. This research has shown quite interesting and promising results using the same classification (SVM) method.

Time-frequency analysis of visual evoked potentials for interhemispheric transfer time and proportion in callosal fibers of different diameters.

This study is an extension of the experimental research of Nalçaci et al., who presented 16 subjects with a reversal of checkerboard pattern as stimuli in the right visual field or left visual field and recorded EEG at O1, O2, P3, and P4. They applied the chosen bandpass filters (4-8, 8-15, 15-20, 20-32 Hz) to the VEPs of subjects and obtained four different components for each VEP. The first aim of this study is to improve the previous report using some methods in time-frequency domain to estimate interhemispheric delays and amplitudes in a time window. Using the improved estimates of interhemispheric delays, the second aim is to estimate the proportion of callosal fibers of different diameters that are activated by visual stimuli by comparing amplitudes of VEPs in different frequency bands. If the relation between frequency components of VEP and delays for callosal fibers of different dimension were reliable, it would give us an opportunity to deal with amplitude of bandpass-filtered VEPs in order to see approximately the proportion of these fibers activated by a certain stimulus. By using frequency-dependent shifts in time and maximizing the cross correlation of direct VEP (DVEP-VEP obtained from contralateral hemisphere)-indirect VEP (IVEP-VEP obtained from ipsilateral hemisphere) pairs in the time-frequency domain, we examined the delay not only at P100 and N160 peaks but along a meaningful time interval as well. Furthermore, by shifting back the IVEP according to the delay estimated at each time window, both the amplitudes and energies of the synchronized DVEP-IVEP pairs were compared at the chosen frequency bands. The percentages of IVEPs at each band was then examined further in conjunction with the distribution of axon diameters in the posterior pole of the CC, questioning the relation between the distributions of the axon diameters and activations at each band. We established an energy definition to express the activation in the fibers. When the energy percentages of IVEPs in theta and alpha were totaled, they were found to be between 76.2% and 81.6%, which is close to the value 74-77% for fibers of 0.4-1 microm in diameter obtained from anatomical study of human CC. The sum of energy percentages in the beta1 and beta2 bands was between 20.1% and 24.2%, which probably reflects the proportion of activation of callosal fibers 1-3 microm in diameter.