[Paper Review] Identification and Visualization of the Underlying Independent Causes of the Diagnostic of Diabetic Retinopathy made by a Deep Learning Classifier

J de la Torre, Aı̈da Valls+2 more|arXiv (Cornell University)|Sep 23, 2018

Retinal Imaging and Analysis1 references3 citations

TL;DR

This paper proposes a method to identify and visualize the independent, mathematically distinct causes a deep learning model uses to classify diabetic retinopathy from fundus images. By applying Independent Component Analysis (ICA) to the final layer features and using score visualization, the authors show that only three independent components are needed to achieve near-human-level classification performance (99% of original metric), with each component highlighting distinct lesion patterns, enabling precise localization of disease signs while filtering out noise and artifacts.

ABSTRACT

Interpretability is a key factor in the design of automatic classifiers for medical diagnosis. Deep learning models have been proven to be a very effective classification algorithm when trained in a supervised way with enough data. The main concern is the difficulty of inferring rationale interpretations from them. Different attempts have been done in last years in order to convert deep learning classifiers from high confidence statistical black box machines into self-explanatory models. In this paper we go forward into the generation of explanations by identifying the independent causes that use a deep learning model for classifying an image into a certain class. We use a combination of Independent Component Analysis with a Score Visualization technique. In this paper we study the medical problem of classifying an eye fundus image into 5 levels of Diabetic Retinopathy. We conclude that only 3 independent components are enough for the differentiation and correct classification between the 5 disease standard classes. We propose a method for visualizing them and detecting lesions from the generated visual maps.

Motivation & Objective

• To address the lack of interpretability in deep learning models for medical diagnosis, particularly in classifying diabetic retinopathy from fundus images.
• To identify the underlying independent statistical regularities that a trained deep learning classifier uses to make diagnostic decisions.
• To develop a visualization method that isolates and localizes these independent causes in the image space for clinical interpretability.
• To determine the minimal number of components required to preserve classification performance, reducing redundancy in feature space.
• To validate the method's ability to distinguish true lesions from noise and artifacts in retinal images.

Proposed method

• The method applies Independent Component Analysis (ICA) to the feature representations from the final fully connected layer of a 17-layer deep neural network trained for diabetic retinopathy classification.
• ICA is used to compress the original 64-dimensional feature space into a minimal set of statistically independent components that retain classification capacity.
• A pixel-wise relevance propagation technique is adapted to visualize the contribution of each ICA component to the final classification decision.
• The visualization method computes score maps by backpropagating relevance through the network, focusing on individual ICA components to localize lesion-related patterns.
• The approach uses a receptive field of 61×61 pixels to compute localized relevance scores, with thresholds set at ±2 standard deviations to highlight significant contributions.
• The method is validated by comparing classification performance (Quadratic Weighted Kappa) and t-SNE visualizations between original and ICA-compressed feature spaces.

Experimental results

Research questions

• RQ1 What is the minimal number of independent components required to preserve the classification performance of a deep learning model for diabetic retinopathy?
• RQ2 Which statistical regularities in the feature space correspond to clinically relevant lesion patterns rather than noise or artifacts?
• RQ3 Can independent component analysis effectively disentangle the underlying causes of classification decisions in a deep learning model?
• RQ4 How accurately can the proposed visualization method localize actual retinal lesions compared to the model’s predictions?
• RQ5 To what extent does the ICA-compressed feature space preserve the model’s diagnostic capability while reducing redundancy?

Key findings

• Only three independent components are sufficient to achieve 99% of the original model’s performance, as measured by Quadratic Weighted Kappa (QWK), with κ_original = 0.800 and κ_ICA = 0.790.
• The ICA-compressed feature space maintains comparable t-SNE visualization quality to the original feature space, indicating preserved structural relationships in the data.
• ICA component 2 shows an almost perfect match with actual retinal lesions, exhibiting very low false positive and false negative rates in score maps.
• ICA components 0 and 1 capture statistical regularities unrelated to lesions, such as noise and artifacts, demonstrating ICA’s ability to filter out non-pathological patterns.
• The pixel-wise score visualization method successfully localizes distinct, mathematically independent disease indicators, enabling interpretable lesion detection.
• The method generalizes well, as the ICA score model is applicable to other medical image classification tasks beyond diabetic retinopathy.