Source-linked AI summary
A Deep Learning Interpretable Classifier for Diabetic Retinopathy Disease Grading
Jordi de la Torre, Aida Valls, Domenec Puig
TL;DR
Deep neural networks perform well in medical image classification but often lack interpretable rationales. This paper develops a diabetic-retinopathy classifier that propagates class scores through network layers to pixels, reports disease severity and visual explanations, and achieves QWKtest 0.844 when combining both eyes.
Problem
Deep learning models can be effective for medical diagnosis, but their millions of parameters make their decisions difficult to interpret, although explanations are important for physicians.
Method
The paper develops a diabetic-retinopathy classifier that assigns class-score contributions across hidden layers and input pixels, including receptive-field contributions and common neural-network blocks.
Results
0.844 QWKtest was achieved by combining features from both eyes with a linear classifier, while the single-eye test value was 0.801.
Takeaways & Limitations
The model classifies five diabetic-retinopathy severity levels and provides per-class pixel-importance maps that human experts can use to interpret predictions.
Takeaways & Limitations
Interpreting the generated maps requires analyzing score maps for every class and different thresholds; the paper leaves the best extraction method for future work.
Abstract
from arXiv · showhide
Deep neural network models have been proven to be very successful in image classification tasks, also for medical diagnosis, but their main concern is its lack of interpretability. They use to work as intuition machines with high statistical confidence but unable to give interpretable explanations about the reported results. The vast amount of parameters of these models make difficult to infer a rationale interpretation from them. In this paper we present a diabetic retinopathy interpretable classifier able to classify retine images into the different levels of disease severity and of explaining its results by assigning a score for every point in the hidden and input space, evaluating its contribution to the final classification in a linear way. The generated visual maps can be interpreted by an expert in order to compare its own knowledge with the interpretation given by the model.
1. Introduction
Deep learning classifiers are effective but difficult to interpret, which is especially important in medical diagnosis. The paper introduces a diabetic-retinopathy classifier that predicts disease severity while assigning pixel- and layer-level scores to explain each decision.
- Deep learning models can classify images effectively, but their millions of parameters make rationale interpretation difficult.
- Medical diagnosis requires explanations so physicians can compare model decisions with their own knowledge.
- The score-propagation approach separates input-dependent contributions from layer receptive-field contributions and supports batch normalization and other neural-network blocks.
- The paper derives score-propagation blocks intended to facilitate transferring the interpretation method to other networks and applications.
- The proposed model classifies retinal images into five standardized diabetic-retinopathy severity classes based on lesion type and number.
- The classifier assigns importance scores to input pixels and supports visual explanations that experts can inspect.
2. Related Work
Related work includes sensitivity maps, layer-wise relevance propagation, and Taylor-type decompositions for interpreting neural-network decisions. The paper reviews their operating principles and identifies constraints such as non-unique relevance splits and dependence on a suitable Taylor root point.
- Prior diabetic-retinopathy classifiers targeted disease grading or severe-case detection, with reported performance approaching or exceeding expert benchmarks.
- Interpretability approaches reviewed include sensitivity maps, layer-wise relevance propagation, and Taylor-type decomposition models.
- Sensitivity maps use input gradients, but gradients describe local function changes without a direct relationship to the explained classification output.
- Layer-wise relevance propagation preserves total relevance across layers and decomposes node relevance into incoming and outgoing messages.
- The relevance-preservation constraints do not ensure a unique or meaningful distribution of the classifier score across nodes.
- Taylor-type decomposition approximates the classification function around a root point selected so that f(I0)=0, representing maximum uncertainty.
- Taylor explanations depend on finding a valid root near the analyzed image because first-order approximation error grows with their Euclidean distance.
3. Receptive field and Pixel-wise Explanation Model
The model decomposes each output score into an input-dependent component and a layer-dependent constant, propagating the former through network operations and mapping the latter to input space through receptive fields.
- Score decomposition: The proposed score model separates each node’s output into input-dependent and constant layer-dependent components.The final score remains a sum of feature-space contribution and subsequent layer contributions, enabling a unique mapping to input space.
- Score propagation: Input-dependent scores are propagated backward through each layer, while constant scores are retained as properties of their receptive fields.The two constituents are treated differently during backward propagation; receptive-field values are mapped to input space afterward.
- Activation nodes: For activation functions, the score is expressed through the activation and can be approximated with a first-order Taylor expansion for general functions.For ReLU family functions, the input-dependent score remains consistent directly; other activations use a nearby expansion point.
- Convolution layers: Convolution scores are distributed across individual inputs, excluding the bias contribution from backpropagation.The bias is treated as an inherent layer property rather than an input-attributable component.
- Pooling and fully connected layers: Pooling and fully connected layers propagate scores according to their structures: selected inputs receive max-pooling scores, equal parts receive average-pooling scores, and linear inputs receive decomposed scores.Fully connected propagation leaves the bias as the layer’s own score contribution.
- Mapping to input space: Hidden-space constant scores are mapped to pixels by summing two-dimensional Gaussian receptive-field distributions, using RF = 2σ as an approximation.The method accounts for effective receptive fields contributing more near their centers than their borders.
4. Classification Model
The classifier combines a convolutional feature extractor with a linear, interpretable prediction stage, using design choices tailored to diabetic-retinopathy grading. Its architecture balances retinal coverage, computational efficiency, training stability, and ordinal-class performance.
- Architecture design: 640-pixel retinal diameter is selected as the optimal resolution because larger tested diameters do not improve results significantly.The study tests retinal diameters from 128 to 892 pixels while considering memory and computation constraints.
- Architecture design: The network uses all available retinal information through a fully convolutional design with stacked 3x3 convolutions and a receptive field matched to the image size.The final receptive field reaches 637x637 for a 640x640 input, while small convolutions reduce parameter requirements and support nonlinear feature learning.
- Training and classification: ReLU activations and batch normalization are used throughout, while QWK is optimized as the loss because the disease grades are ordinal.The paper reports that batch normalization facilitates training and that directly optimizing QWK gives better classification results.
- Training and classification: A linear final classifier is chosen for simplicity and interpretability, with softmax converting class scores into probabilities.The model can determine the predicted class from the largest score without explicitly evaluating softmax.
- Architecture design: The 391,325-parameter CNN has 17 layers, a feature-extraction stage, and a classifier that produces a 64-feature vector before linear class scoring.The network uses 2x2 max-pooling between feature blocks, followed by a 2x2 convolution and 4x4 average-pooling.
5. Results
The classifier reaches strong DR grading performance while producing class-specific scores and pixel-level maps that expose how predictions are formed. These maps support expert inspection of image evidence and potential analysis of unexpected classifications.
- Classification: 0.814 QWK on validation and 0.801 on the unseen test set; combining both eyes increases QWKtest to 0.844.The reported expert ophthalmologist inter-rating agreement is in the 0.80s.
- Classification: The model retains the five disease grades so it can encode features separating intermediate classes, accepting reduced accuracy to preserve interpretability.Aggregated detection could increase accuracy but would miss features distinguishing positive classes 1 to 4.
- Explanations: For a class-4 test image, C4 = +167.1 is the highest pre-soft-max score, so the image is correctly classified as class 4.The other class scores are C0 = −638.9, C1 = −379.7, C2 = −114.6, and C3 = +62.8.
- Explanations: Layer-16 feature scores differ across predicted classes, showing that the same learned features contribute differently to class decisions.The reported visualization presents individual feature scores for layer 16.
- Explanations: Class-specific score maps distribute the final inference score across pixels, making each pixel’s relative contribution to the decision visible.The pixel-score sum equals the last-layer inference score, and visualizations can be restricted to positive or thresholded scores.
- Explanations: A complete analysis should compare score maps for every class and use different thresholds, but the figure presents only the predicted-class map.The authors identify extracting conclusions from these maps as future work.
6. Conclusions
The paper introduces an exact score-propagation framework for explaining deep-learning classifications and applies it to a five-level diabetic-retinopathy classifier. The resulting class-specific pixel-importance maps provide human experts with information for interpreting model decisions.
- Conclusions: The method distributes last-layer scores across input pixels and derives score propagation for typical deep-learning building blocks.It uses a modified pixel-wise relevance propagation method that back-propagates input-dependent scores while retaining constant layer contributions.
- Conclusions: The DR classifier grades retinal images across five standardized severity levels and reports pixel-importance maps for every class.These maps are intended to support human inference and interpretation.