Interpretable Deep Learning Model Research Articles

ICVM: An Interpretable Deep Learning Model for CVM Assessment Under Label Uncertainty.

The Cervical Vertebral Maturation (CVM) method aims to determine the craniofacial skeletal maturational stage, which is crucial for orthodontic and orthopedic treatment. In this paper, we explore the potential of deep learning for automatic CVM assessment. In particular, we propose a convolutional neural network named iCVM. Based on the residual network, it is specialized for the challenges unique to the task of CVM assessment. 1) To combat overfitting due to limited data size, multiple dropout layers are utilized. 2) To address the inevitable label ambiguity between adjacent maturational stages, we introduce the concept of label distribution learning in the loss function. Besides, we attempt to analyze the regions important for the prediction of the model by using the Grad-CAM technique. The learned strategy shows surprisingly high consistency with the clinical criteria. This indicates that the decisions made by our model are well interpretable, which is critical in evaluation of growth and development in orthodontics. Moreover, to drive future research in the field, we release a new dataset named CVM-900 along with the paper. It contains the cervical part of 900 lateral cephalograms collected from orthodontic patients of different ages and genders. Experimental results show that the proposed approach achieves superior performance on CVM-900 in terms of various evaluation metrics.

Towards the interpretability of deep learning models for multi-modal neuroimaging: Finding structural changes of the ageing brain

Brain-age (BA) estimates based on deep learning are increasingly used as neuroimaging biomarker for brain health; however, the underlying neural features have remained unclear. We combined ensembles of convolutional neural networks with Layer-wise Relevance Propagation (LRP) to detect which brain features contribute to BA. Trained on magnetic resonance imaging (MRI) data of a population-based study (n = 2637, 18–82 years), our models estimated age accurately based on single and multiple modalities, regionally restricted and whole-brain images (mean absolute errors 3.37–3.86 years). We find that BA estimates capture ageing at both small and large-scale changes, revealing gross enlargements of ventricles and subarachnoid spaces, as well as white matter lesions, and atrophies that appear throughout the brain. Divergence from expected ageing reflected cardiovascular risk factors and accelerated ageing was more pronounced in the frontal lobe. Applying LRP, our study demonstrates how superior deep learning models detect brain-ageing in healthy and at-risk individuals throughout adulthood.

This Looks Like That There: Interpretable Neural Networks for Image Tasks When Location Matters

Abstract We develop and demonstrate a new interpretable deep learning model specifically designed for image analysis in Earth system science applications. The neural network is designed to be inherently interpretable, rather than explained via post hoc methods. This is achieved by training the network to identify parts of training images that act as prototypes for correctly classifying unseen images. The new network architecture extends the interpretable prototype architecture of a previous study in computer science to incorporate absolute location. This is useful for Earth system science where images are typically the result of physics-based processes, and the information is often geolocated. Although the network is constrained to only learn via similarities to a small number of learned prototypes, it can be trained to exhibit only a minimal reduction in accuracy relative to noninterpretable architectures. We apply the new model to two Earth science use cases: a synthetic dataset that loosely represents atmospheric high and low pressure systems, and atmospheric reanalysis fields to identify the state of tropical convective activity associated with the Madden–Julian oscillation. In both cases, we demonstrate that considering absolute location greatly improves testing accuracies when compared with a location-agnostic method. Furthermore, the network architecture identifies specific historical dates that capture multivariate, prototypical behavior of tropical climate variability. Significance Statement Machine learning models are incredibly powerful predictors but are often opaque “black boxes.” The how-and-why the model makes its predictions is inscrutable—the model is not interpretable. We introduce a new machine learning model specifically designed for image analysis in Earth system science applications. The model is designed to be inherently interpretable and extends previous work in computer science to incorporate location information. This is important because images in Earth system science are typically the result of physics-based processes, and the information is often map based. We demonstrate its use for two Earth science use cases and show that the interpretable network exhibits only a small reduction in accuracy relative to black-box models.

BayeshERG: a robust, reliable and interpretable deep learning model for predicting hERG channel blockers.

Unintended inhibition of the human ether-à-go-go-related gene (hERG) ion channel by small molecules leads to severe cardiotoxicity. Thus, hERG channel blockage is a significant concern in the development of new drugs. Several computational models have been developed to predict hERG channel blockage, including deep learning models; however, they lack robustness, reliability and interpretability. Here, we developed a graph-based Bayesian deep learning model for hERG channel blocker prediction, named BayeshERG, which has robust predictive power, high reliability and high resolution of interpretability. First, we applied transfer learning with 300000 large data in initial pre-training to increase the predictive performance. Second, we implemented a Bayesian neural network with Monte Carlo dropout to calibrate the uncertainty of the prediction. Third, we utilized global multihead attentive pooling to augment the high resolution of structural interpretability for the hERG channel blockers and nonblockers. We conducted both internal and external validations for stringent evaluation; in particular, we benchmarked most of the publicly available hERG channel blocker prediction models. We showed that our proposed model outperformed predictive performance and uncertainty calibration performance. Furthermore, we found that our model learned to focus on the essential substructures of hERG channel blockers via an attention mechanism. Finally, we validated the prediction results of our model by conducting in vitro experiments and confirmed its high validity. In summary, BayeshERG could serve as a versatile tool for discovering hERG channel blockers and helping maximize the possibility of successful drug discovery. The data and source code are available at our GitHub repository (https://github.com/GIST-CSBL/BayeshERG).

Abstract 636: Prediction of therapeutic response via data-driven maps of tumor cell architecture

Abstract Machine learning is revolutionizing the analysis of biological and biomedical data. One continued challenge is to build models that are not only predictive but also provide mechanistic explanations. Directly incorporating explainability into the model’s design has emerged as a possible solution. Recently, we reported several “visible” machine learning systems - DCell and DrugCell - which predict cellular growth based on genetic perturbations. In these visible models, the structure of the underlying neural network mirrors known aspects of cell biology. Here we designed a cancer-specific model of therapeutic response. First, we selected NeST (Nested Systems in Tumors; Zheng et al., Science 2021), a data-driven hierarchy of tumor cell systems under selection in cancer, as the structure for our visible neural network. Second, we restricted input features to genes currently assayed on clinical cancer gene panels. We considered training schemes that assess the model’s ability to generalize to novel genotypes. Most importantly, we assessed performance on patient data from the AACR Project GENIE clinical trial where we discriminated between patients who are likely to respond to selected drug treatments. Finally, we examined model explanations within the visible NeST architecture to highlight cellular mechanisms of response to drug therapy. Overall, we found that by considering a tumor cell architecture optimized for clinical application, we can design an interpretable deep learning model that both accurately stratifies patient responses and explains the drug mechanism in cancer cells. Citation Format: Erica N. Silva, Akshat Singhal, Sungjoon Park, Jason Kreisberg, Trey Ideker. Prediction of therapeutic response via data-driven maps of tumor cell architecture [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 636.

Interpretable deep learning models for hourly solar radiation prediction based on graph neural network and attention

With the rapid development of high-performance computing technology, data-driven models, especially deep learning models, are being used increasingly for solar radiation prediction. However, the characteristics of the black box model lead to a lack of interpretability in their prediction results. This limits the application of the model in final optimization scenarios (such as model predictive control), as operation managers might not fully trust models lacking explanatory results. In our study, models were proposed based on the prediction model of the recurrent neural network. We hope to improve the interpretability of the models through the design and improvement of the model structure, thereby increasing the credibility of the model results. The interpretability in time and spatial dependencies of the prediction process were studied by the attention mechanism and graph neural network, respectively. Our results showed that the deep learning model, with attention, could effectively shift the attention mechanism to adapt to varying prediction target hours. The graph neural network expresses the most relevant variables in the dataset related to solar radiation through a self-learning graph structure. The results showed that solar radiation is connected directly with month, hour, temperature, penetrating rainfall, water vapor pressure, and radiation time.

Decoding neural activity preceding balance loss during standing with a lower-limb exoskeleton using an interpretable deep learning model

Objective: Falls are a leading cause of death in adults 65 and older. Recent efforts to restore lower-limb function in these populations have seen an increase in the use of wearable robotic systems; however, fall prevention measures in these systems require early detection of balance loss to be effective. Prior studies have investigated whether kinematic variables contain information about an impending fall, but few have examined the potential of using electroencephalography (EEG) as a fall-predicting signal and how the brain responds to avoid a fall. Approach: To address this, we decoded neural activity in a balance perturbation task while wearing an exoskeleton. We acquired EEG, electromyography (EMG), and center of pressure (COP) data from seven healthy participants during mechanical perturbations while standing. The timing of the perturbations was randomized in all trials. Main results: We found perturbation evoked potentials (PEP) components as early as 75–134 ms after the onset of the external perturbation, which preceded both the peak in EMG (∼180 ms) and the COP (∼350 ms). A convolutional neural network trained to predict balance perturbations from single-trial EEG had a mean F-score of 75.0 ± 4.3. Clustering GradCAM-based model explanations demonstrated that the model utilized components in the PEP and was not driven by artifacts. Additionally, dynamic functional connectivity results agreed with model explanations; the nodal connectivity measured using phase difference derivative was higher in the occipital-parietal region in the early stage of perturbations, before shifting to the parietal, motor, and back to the frontal-parietal channels. Continuous-time decoding of COP trajectories from EEG, using a gated recurrent unit model, achieved a mean Pearson’s correlation coefficient of 0.7 ± 0.06. Significance: Overall, our findings suggest that EEG signals contain short-latency neural information related to an impending fall, which may be useful for developing brain-machine interface systems for fall prevention in robotic exoskeletons.

IDL-PPBopt: A Strategy for Prediction and Optimization of Human Plasma Protein Binding of Compounds via an Interpretable Deep Learning Method.

The prediction and optimization of pharmacokinetic properties are essential in lead optimization. Traditional strategies mainly depend on the empirical chemical rules from medicinal chemists. However, with the rising amount of data, it is getting more difficult to manually extract useful medicinal chemistry knowledge. To this end, we introduced IDL-PPBopt, a computational strategy for predicting and optimizing the plasma protein binding (PPB) property based on an interpretable deep learning method. At first, a curated PPB data set was used to construct an interpretable deep learning model, which showed excellent predictive performance with a root mean squared error of 0.112 for the entire test set. Then, we designed a detection protocol based on the model and Wilcoxon test to identify the PPB-related substructures (named privileged substructures, PSubs) for each molecule. In total, 22 general privileged substructures (GPSubs) were identified, which shared some common features such as nitrogen-containing groups, diamines with two carbon units, and azetidine. Furthermore, a series of second-level chemical rules for each GPSub were derived through a statistical test and then summarized into substructure pairs. We demonstrated that these substructure pairs were equally applicable outside the training set and accordingly customized the structural modification schemes for each GPSub, which provided alternatives for the optimization of the PPB property. Therefore, IDL-PPBopt provides a promising scheme for the prediction and optimization of the PPB property and would be helpful for lead optimization of other pharmacokinetic properties.

Interpretable Deep Learning Model Reveals Subsequences of Various Functions for Long Non-Coding RNA Identification.

Long non-coding RNAs (lncRNAs) play crucial roles in many biological processes and are implicated in several diseases. With the next-generation sequencing technologies, substantial unannotated transcripts have been discovered. Classifying unannotated transcripts using biological experiments are more time-consuming and expensive than computational approaches. Several tools are available for identifying long non-coding RNAs. These tools, however, did not explain the features in their tools that contributed to the prediction results. Here, we present Xlnc1DCNN, a tool for distinguishing long non-coding RNAs (lncRNAs) from protein-coding transcripts (PCTs) using a one-dimensional convolutional neural network with prediction explanations. The evaluation results of the human test set showed that Xlnc1DCNN outperformed other state-of-the-art tools in terms of accuracy and F1-score. The explanation results revealed that lncRNA transcripts were mainly identified as sequences with no conserved regions, short patterns with unknown functions, or only regions of transmembrane helices while protein-coding transcripts were mostly classified by conserved protein domains or families. The explanation results also conveyed the probably inconsistent annotations among the public databases, lncRNA transcripts which contain protein domains, protein families, or intrinsically disordered regions (IDRs). Xlnc1DCNN is freely available at https://github.com/cucpbioinfo/Xlnc1DCNN.

Reliable quality assurance of X-ray mammography scanner by evaluation the standard mammography phantom image using an interpretable deep learning model

ObjectiveMammography is the initial examination to detect breast cancer symptoms, and quality control of mammography devices is crucial to maintain accurate diagnosis and to safeguard against degradation of performance. The objective of this study was to assist radiologists in mammography phantom image evaluation by developing and validating an interpretable deep learning model capable of objectively evaluating the quality of standard phantom images for mammography. Materials and MethodsA total of 2,208 mammography phantom images were collected for periodic accreditation of the scanner from 1,755 institutions. The dataset was randomly split into training (1,808 images) and testing (400 images) datasets with subgroups (76 images) for the multi-reader study. To develop an interpretable model that contains two deep learning networks in series, five processing steps were performed: mammography phantom detection, phantom object detection, post-processing, score evaluation, and a report with a comment about ambiguous results. ResultsFor phantom detection, the accuracy and mean intersection over union (mIOU) were 1.00 and 0.938 in the test dataset, respectively. During phantom object detection, a total of 6,369 out of 6,400 objects were detected as the correct object class, and the accuracy and mIOU were 0.995 and 0.813, respectively. The predicted score for each object showed a consensus of 97.40% excluding ambiguous points and 59.10% for ambiguous points of the groups. ConclusionsThe interpretable deep learning model using large-scale data from multiple centers shows high performance and reasonable object scoring, successfully validating the reliability and feasibility of mammography phantom image quality management.

Estimation of Atmospheric PM10 Concentration in China Using an Interpretable Deep Learning Model and Top‐of‐the‐Atmosphere Reflectance Data From China’s New Generation Geostationary Meteorological Satellite, FY‐4A

AbstractThe rapid urbanization in China and the long‐range transport dust (LRTD) from arid and semi‐arid areas has resulted in an increase of PM10 concentration. In this study, an interpretable deep learning model [deep forest (DF)] with FY‐4A top‐of‐the‐atmosphere reflectance (TOAR) data were used to obtain the hourly PM10 in China. The optimal hourly average R2 of 10‐fold cross validation can achieve 0.85 (13:00 Beijing time); The R2 (RMSE, μg/m³) of the daily, monthly, and annual averages were 0.82 (24.16), 0.97 (6.53), and 0.99 (2.30), respectively. Using TOAR data, the DF model performed better than other machine learning models. The feature importance of the TOAR‐PM10 model showed that TOAR and meteorological elements both contributed significantly to the model. In spring, the PM10 in northern China was greater than that in southern China, which may be related to the LRTD. Excluding the dust weather periods, the areas with high PM10 values in China were mainly in cities and their suburbs, where were correlated with human activities. During a dust weather process, LRTD increased PM10 in northern China by 80.4%. During a mixture haze and dust weather process, the PM10 increased by 130.2% in northern China, of which LRTD led to an increase of 73.7%. The sources (from the Taklimakan Desert in China) and transmission paths of these two LRTD processes were similar. The contribution of LRTD to PM10 was related to dust intensity and meteorological conditions. The results showed that LRTD and local pollution to PM10 was both important in haze periods.

Explainable Deep Learning Study for Leaf Disease Classification

Explainable artificial intelligence has been extensively studied recently. However, the research of interpretable methods in the agricultural field has not been systematically studied. We studied the interpretability of deep learning models in different agricultural classification tasks based on the fruit leaves dataset. The purpose is to explore whether the classification model is more inclined to extract the appearance characteristics of leaves or the texture characteristics of leaf lesions during the feature extraction process. The dataset was arranged into three experiments with different categories. In each experiment, the VGG, GoogLeNet, and ResNet models were used and the ResNet-attention model was applied with three interpretable methods. The results show that the ResNet model has the highest accuracy rate in the three experiments, which are 99.11%, 99.4%, and 99.89%, respectively. It is also found that the attention module could improve the feature extraction of the model, and clarify the focus of the model in different experiments when extracting features. These results will help agricultural practitioners better apply deep learning models to solve more practical problems.

ISeeU2: Visually interpretable mortality prediction inside the ICU using deep learning and free-text medical notes

Accurate mortality prediction allows Intensive Care Units (ICUs) to adequately benchmark clinical practice and identify patients with unexpected outcomes. Traditionally, simple statistical models have been used to assess patient death risk, many times with sub-optimal performance. On the other hand Deep Learning holds promise to positively impact clinical practice by leveraging medical data to assist diagnosis and prediction, including mortality prediction. However, as the question of whether powerful Deep Learning models attend correlations backed by sound medical knowledge when generating predictions remains open, additional interpretability tools are needed to foster trust and encourage the use of AI by clinicians. In this work we show an interpretable Deep Learning model trained on MIMIC-III to predict mortality inside the ICU using raw nursing notes, together with visual explanations for word importance based on the Shapley Value. Our model reaches a ROC of 0.8629 (±0.0058), outperforming the traditional SAPS-II score and a LSTM recurrent neural network baseline while providing enhanced interpretability when compared with similar Deep Learning approaches. Supporting code can be found at https://github.com/williamcaicedo/ISeeU2.

Artificial intelligence assessment for early detection and prediction of renal impairment using electrocardiography

PurposeAlthough renal failure is a major healthcare burden globally and the cornerstone for preventing its irreversible progression is an early diagnosis, an adequate and noninvasive tool to screen renal impairment (RI) reliably and economically does not exist. We developed an interpretable deep learning model (DLM) using electrocardiography (ECG) and validated its performance.MethodsThis retrospective cohort study included two hospitals. We included 115,361 patients who had at least one ECG taken with an estimated glomerular filtration rate measurement within 30 min of the index ECG. A DLM was developed using 96,549 ECGs of 55,222 patients. The internal validation included 22,949 ECGs of 22,949 patients. Furthermore, we conducted an external validation with 37,190 ECGs of 37,190 patients from another hospital. The endpoint was to detect a moderate to severe RI (estimated glomerular filtration rate < 45 ml/min/1.73m2).ResultsThe area under the receiver operating characteristic curve (AUC) of a DLM using a 12-lead ECG for detecting RI during the internal and external validation was 0.858 (95% confidence interval 0.851–0.866) and 0.906 (0.900–0.912), respectively. In the initial evaluation of 25,536 individuals without RI patients whose DLM was defined as having a higher risk had a significantly higher chance of developing RI than those in the low-risk group (17.2% vs. 2.4%, p < 0.001). The sensitivity map indicated that the DLM focused on the QRS complex and T-wave for detecting RI.ConclusionThe DLM demonstrated high performance for RI detection and prediction using 12-, 6-, single-lead ECGs.

Emerging artificial intelligence methods for fighting lung cancer: A survey

Lung cancer has one of the highest incidence rates and mortality rates among all common cancers worldwide. Early detection of suspicious lung nodules is crucial in fighting lung cancer. In recent years, with the proliferation of clinical data like low-dose computed tomography (LDCT), histology whole slide images, electronic health records, and sensor readings from medical IoT devices etc., many artificial intelligence tools have taken more important roles in lung cancer management. In this survey, we lay out the current and emergent artificial intelligence methods for fighting lung cancers. Besides the commonly used CT image based deep learning models for detecting and diagnosing lung nodules, we also cover emergent AI techniques for lung cancer: 1) federated deep learning models for harnessing multi-center data with privacy in mind, 2) multi-modal deep learning models for integrating multiple sources of clinical and image data, 3) interpretable deep learning models for opening the black box for clinicians. In the big data era for cancer management, we believe this short survey will help AI researchers better understand the clinical challenges of lung cancer and will also help clinicians better understand the emergent AI tools.

Think positive: An interpretable neural network for image recognition

The COVID-19 pandemic is an ongoing pandemic and is placing additional burden on healthcare systems around the world. Timely and effectively detecting the virus can help to reduce the spread of the disease. Although, RT-PCR is still a gold standard for COVID-19 testing, deep learning models to identify the virus from medical images can also be helpful in certain circumstances. In particular, in situations when patients undergo routine X-rays and/or CT-scans tests but within a few days of such tests they develop respiratory complications. Deep learning models can also be used for pre-screening prior to RT-PCR testing. However, the transparency/interpretability of the reasoning process of predictions made by such deep learning models is essential. In this paper, we propose an interpretable deep learning model that uses positive reasoning process to make predictions. We trained and tested our model over the dataset of chest CT-scan images of COVID-19 patients, normal people and pneumonia patients. Our model gives the accuracy, precision, recall and F-score equal to 99.48%, 0.99, 0.99 and 0.99, respectively.

Interpretability Analysis of One-Year Mortality Prediction for Stroke Patients Based on Deep Neural Network.

Clinically, physicians collect the benchmark medical data to establish archives for a stroke patient and then add the follow up data regularly. It has great significance on prognosis prediction for stroke patients. In this paper, we present an interpretable deep learning model to predict the one-year mortality risk on stroke. We design sub-modules to reconstruct features from original clinical data that highlight the dissimilarity and temporality of different variables. The model consists of Bidirectional Long Short-Term Memory (Bi-LSTM), in which a novel correlation attention module is proposed that takes the correlation of variables into consideration. In experiments, datasets are collected clinically from the department of neurology in a local AAA hospital. It consists of 2,275 stroke patients hospitalized in the department of neurology from 2014 to 2016. Our model achieves a precision of 0.9414, a recall of 0.9502 and an F1-score of 0.9415. In addition, we provide the analysis of the interpretability by visualizations with reference to clinical professional guidelines.

Exploring kinase family inhibitors and their moiety preferences using deep SHapley additive exPlanations

BackgroundWhile it has been known that human protein kinases mediate most signal transductions in cells and their dysfunction can result in inflammatory diseases and cancers, it remains a challenge to find effective kinase inhibitor as drugs for these diseases. One major challenge is the compensatory upregulation of related kinases following some critical kinase inhibition. To circumvent the compensatory effect, it is desirable to have inhibitors that inhibit all the kinases belonging to the same family, instead of targeting only a few kinases. However, finding inhibitors that target a whole kinase family is laborious and time consuming in wet lab.ResultsIn this paper, we present a computational approach taking advantage of interpretable deep learning models to address this challenge. Specifically, we firstly collected 9,037 inhibitor bioassay results (with 3991 active and 5046 inactive pairs) for eight kinase families (including EGFR, Jak, GSK, CLK, PIM, PKD, Akt and PKG) from the ChEMBL25 Database and the Metz Kinase Profiling Data. We generated 238 binary moiety features for each inhibitor, and used the features as input to train eight deep neural networks (DNN) models to predict whether an inhibitor is active for each kinase family. We then employed the SHapley Additive exPlanations (SHAP) to analyze the importance of each moiety feature in each classification model, identifying moieties that are in the common kinase hinge sites across the eight kinase families, as well as moieties that are specific to some kinase families. We finally validated these identified moieties using experimental crystal structures to reveal their functional importance in kinase inhibition.ConclusionWith the SHAP methodology, we identified two common moieties for eight kinase families, 9 EGFR-specific moieties, and 6 Akt-specific moieties, that bear functional importance in kinase inhibition. Our result suggests that SHAP has the potential to help finding effective pan-kinase family inhibitors.

Interpretable deep learning for the prediction of ICU admission likelihood and mortality of COVID-19 patients.

The global healthcare system is being overburdened by an increasing number of COVID-19 patients. Physicians are having difficulty allocating resources and focusing their attention on high-risk patients, partly due to the difficulty in identifying high-risk patients early. COVID-19 hospitalizations require specialized treatment capabilities and can cause a burden on healthcare resources. Estimating future hospitalization of COVID-19 patients is, therefore, crucial to saving lives. In this paper, an interpretable deep learning model is developed to predict intensive care unit (ICU) admission and mortality of COVID-19 patients. The study comprised of patients from the Stony Brook University Hospital, with patient information such as demographics, comorbidities, symptoms, vital signs, and laboratory tests recorded. The top three predictors of ICU admission were ferritin, diarrhoea, and alamine aminotransferase, and the top predictors for mortality were COPD, ferritin, and myalgia. The proposed model predicted ICU admission with an AUC score of 88.3% and predicted mortality with an AUC score of 96.3%. The proposed model was evaluated against existing model in the literature which achieved an AUC of 72.8% in predicting ICU admission and achieved an AUC of 84.4% in predicting mortality. It can clearly be seen that the model proposed in this paper shows superiority over existing models. The proposed model has the potential to provide tools to frontline doctors to help classify patients in time-bound and resource-limited scenarios.

Emergence of Direction-Selective Retinal Cell Types in Task-Optimized Deep Learning Models.

Convolutional neural networks (CNNs), a class of deep learning models, have experienced recent success in modeling sensory cortices and retinal circuits through optimizing performance on machine learning tasks, otherwise known as task optimization. Previous research has shown task-optimized CNNs to be capable of providing explanations as to why the retina efficiently encodes natural stimuli and how certain retinal cell types are involved in efficient encoding. In our work, we sought to use task-optimized CNNs as a means of explaining computational mechanisms responsible for motion-selective retinal circuits. We designed a biologically constrained CNN and optimized its performance on a motion-classification task. We drew inspiration from psychophysics, deep learning, and systems neuroscience literature to develop a toolbox of methods to reverse engineer the computational mechanisms learned in our model. Through reverse engineering our model, we proposed a computational mechanism in which direction-selective ganglion cells and starburst amacrine cells, both experimentally observed retinal cell types, emerge in our model to discriminate among moving stimuli. This emergence suggests that direction-selective circuits in the retina are ecologically designed to robustly discriminate among moving stimuli. Our results and methods also provide a framework for how to build more interpretable deep learning models and how to understand them.