A Deep Learning Interpretable Classifier for Diabetic Retinopathy Disease Grading (1712.08107v1)

Abstract: 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.

• The paper presents an interpretable deep learning classifier that decomposes scores into pixel-based and constant components for DR grading.
• It employs innovative score propagation techniques to generate visual explanations that aid clinical validation of retinal diagnoses.
• Tested on EyePACS, the model attains a quadrature weighted kappa up to 0.844, showcasing expert-level diagnostic performance.

An Interpretable Deep Learning Model for Diabetic Retinopathy Disease Grading

The deep learning community continues to address the challenge of interpretable models, especially in critical domains such as medical diagnosis. In this context, the paper "A Deep Learning Interpretable Classifier for Diabetic Retinopathy Disease Grading" by Jordi de la Torre, Aida Valls, and Domenec Puig presents a method for making the output of a deep neural network interpretable by practitioners in the domain of diabetic retinopathy (DR) classification.

Overview of the Approach

The authors propose an innovative approach using a deep learning model that not only classifies retinal images according to the severity of DR but also provides interpretable visual explanations of the classification decisions. The method extends previous work on pixel-wise relevance propagation by introducing an explicit decomposition of the classification score into a pixel-based explanation and a constant component tied to the architecture.

The model is built to identify five stages of DR severity, ranging from no apparent retinopathy to proliferative DR, by leveraging retinal images from the EyePACS dataset. The unique feature of this approach lies in its capability to attribute classification decisions to specific input pixels and thereby provide an "importance score map" for each image. This feature allows clinicians to cross-verify the model's conclusions with their domain knowledge.

Methodological Innovations

The core methodological contribution lies in the reformulation of score propagation techniques within deep learning architectures. Specifically, the model departs from traditional layer-wise conservation of relevance scores and introduces layer-distributed constants. These constants represent contributions from the layers that affect classification scores independently of input pixel values.

This approach recognized two distinct components within the score of each neuron:

1. An input-dependent score reflecting traditional contributions based on input activations.
2. A constant score emanating from the architectural configuration, which the authors map back to the input space using receptive fields assumed to have Gaussian distributions.

Such methodologies offer a more granular interpretation compared to existing models, such as layer-wise relevance propagation (LRP) or Deep Taylor decomposition, by providing a more exact mapping of scores from hidden layers to input pixels.

Numerical Results

Applied to the EyePACS dataset, the model demonstrated prediction performance comparable to human expert level. The classifier scored a quadrature weighted kappa (QWK) of 0.814 on the validation set and improved to 0.844 upon aggregating inputs from both eyes, suggesting eyebrow-raising performance in terms of sensitivity and specificity—key metrics for medical diagnostics.

Theoretical and Practical Implications

Theoretically, the paper offers a substantial contribution to interpretability in AI, particularly for convolutional neural networks. By tackling one of the core limitations of deep learning models, the paper provides a potential pathway for balancing the complexity of deep learning models with the need for transparency and reliability.

Practically, these advancements hold significant promise for clinical decision support systems. With reliable interpretability, models can serve as an extension of the medical professional's diagnostic toolkit, enhancing the availability and accuracy of DR diagnostics while also potentially reducing the workload on retinal experts.

Future Prospects

The implications of this work extend beyond diabetic retinopathy and signal further developments in more transparent AI models. Anticipated future research could focus on refining this interpretability framework for other medical applications, or domains requiring high-stakes decision-making processes. Moreover, improving the computational efficiency of propagating scores across complex architectures may enhance the scalability of this approach.

In conclusion, this paper represents a meaningful stride towards harmonizing the power of deep learning with domain-relevant interpretability mechanisms, offering a template for future exploration and validation across various AI application domains.