Deeply Explainable Artificial Neural Network (2505.06731v1)

Abstract: While deep learning models have demonstrated remarkable success in numerous domains, their black-box nature remains a significant limitation, especially in critical fields such as medical image analysis and inference. Existing explainability methods, such as SHAP, LIME, and Grad-CAM, are typically applied post hoc, adding computational overhead and sometimes producing inconsistent or ambiguous results. In this paper, we present the Deeply Explainable Artificial Neural Network (DxANN), a novel deep learning architecture that embeds explainability ante hoc, directly into the training process. Unlike conventional models that require external interpretation methods, DxANN is designed to produce per-sample, per-feature explanations as part of the forward pass. Built on a flow-based framework, it enables both accurate predictions and transparent decision-making, and is particularly well-suited for image-based tasks. While our focus is on medical imaging, the DxANN architecture is readily adaptable to other data modalities, including tabular and sequential data. DxANN marks a step forward toward intrinsically interpretable deep learning, offering a practical solution for applications where trust and accountability are essential.

Deeply Explainable Artificial Neural Network: Ante Hoc Interpretability in Medical Imaging

The paper introduces the Deeply Explainable Artificial Neural Network (DxANN), a novel architecture designed to incorporate explainability within the model's training process, rather than relying on post hoc interpretative tools. This approach addresses the critical demand for transparency in artificial neural networks, particularly within domains such as medicine where understanding diagnostic decisions is crucial.

Deep learning has made substantial contributions to various fields, yet the opacity of traditional neural networks poses significant challenges in critical applications requiring trust and accountability. Conventional explainability methods, such as SHAP, LIME, and Grad-CAM, are applied after the models are trained, often leading to inconsistent interpretations and additional computational burdens. DxANN, conversely, integrates interpretability during the model training through the Explainability Contribution Score (ECS), offering robust and consistent per-sample, per-feature explanations.

Architectural Framework and Theoretical Foundations

The core innovation of DxANN is its architectural design based on the Real-valued Non-Volume Preserving (Real-NVP) algorithm. This flow-based generative model maps complex data distributions to simplified forms through bijective transformations defined by learnable parameters. DxANN utilizes the Real-NVP framework to develop a binary classifier, with ECS providing a metric for feature importance. ECS is proportional to the distance between the embedded feature in the latent space and the mean of its assigned class distribution, thus calculating the contribution each feature makes to the model’s inference.

This mechanism ensures that DxANN not only achieves high predictions accuracy but also generates interpretative insights that align with domain-specific requirements. For image data, each pixel's significance is assessed in correlation with its spatial context, with ECS quantifying how challenging it is for the model to reconcile a pixel with its class-based embedding representation.

Experimental Validation and Results

The empirical evaluation of DxANN involved two distinct medical image datasets: OCT retinal images and knee X-ray images, each corresponding to the diagnosis of Diabetic Macular Edema (DME) and osteoarthritis (OA), respectively. The datasets were divided into training and test segments, allowing DxANN to be trained comprehensively. Comparative analyses with state-of-the-art models—ResNet-50, VGG-16, and custom CNN architectures—demonstrated DxANN's competitive performance in terms of accuracy.

The DxANN achieved a DME diagnosis accuracy of 97.1% and OA diagnosis accuracy of 97.2%, comparable to and sometimes surpassing the foundation models. Moreover, the ECS-based heatmaps produced by DxANN offered explicit visual explanations that corresponded closely with standard clinical insights, highlighting pathologically relevant regions and structural areas pivotal for diagnosis.

Implications and Future Directions

DxANN represents a significant progression in intrinsically interpretable deep learning by embedding explainability directly within model architecture, thereby eliminating the limitations and inconsistencies of post hoc methods. The broader implications of this work include enhanced trust in AI applications across medically-focused domains and streamlined regulatory approval processes.

Future research could explore the applicability of DxANN to varied data modalities beyond images, integrating sequential and tabular data, and investigating its implementation in other neural architectures such as MLPs, transformers, and RNNs. Additionally, expanding the framework to support multimodal inputs could further bolster its utility in complex diagnostic tasks, paving the way for more holistic AI-driven solutions.

Overall, the paper provides a comprehensive account of DxANN's theoretical foundations, architectural innovations, and empirical validations, underscoring the potential for ante hoc explainability to transform AI applications in critical sectors.