Justifying diagnosis decisions by deep neural networks

Graham Spinks, Marie-Francine Moens

This paper, titled "Justifying Diagnosis Decisions by Deep Neural Networks," by Graham Spinks and Marie-Francine Moens, presents a novel framework for making medical AI more transparent. Instead of just providing a label, the system generates a textual explanation and a "counterfactual" image to show what the patient's X-ray would look like if they had a different diagnosis.

TL;DR

Researchers have developed a multi-task neural network that justifies its medical diagnoses using two distinct methods: generating descriptive text and creating "counterfactual" X-ray images. By showing clinicians what the "nearest alternative" diagnosis would look like for a specific patient, the system provides much clearer evidence than traditional heatmaps.

What Problem Does This Paper Address?

Deep learning models are increasingly accurate at diagnosing diseases from medical images, but they remain "black boxes." In a clinical setting, a doctor cannot simply trust a "90% probability" without knowing why the model reached that conclusion.

The current standard for "explainable AI" in medical imaging is the saliency map (or heatmap), which highlights pixels the model focused on. However, these maps are often noisy and ambiguous; they show where the model looked, but not what it saw or how those features relate to the disease. The authors aim to bridge this gap by providing justifications that are both human-readable (text) and visually comparative (counterfactual images).

Methods

The authors propose a holistic system that operates across three stages:

  1. Textual Embedding (ARAE): They use an Adversarially Regularized Autoencoder (ARAE) to map medical reports (diagnosis + findings) into a smooth, continuous latent space. This ensures that similar medical descriptions are physically close to each other in the model's "mind."
  2. Text-to-Image GAN: Using a modified StackGAN architecture, the model learns to generate realistic X-ray images based on these textual embeddings.
  3. Inverse Mapping ($I$): This is the core innovation. The researchers train a convolutional network to take a real X-ray and map it back into the smooth latent space created in step 1.

The Justification Mechanism: * Textual: The model decodes the latent vector into a diagnosis and a caption (e.g., "Cardiomegaly: heart size slightly elevated"). * Visual (Mapping $M$): To provide a visual justification, the system takes the patient's latent vector and finds the closest point in the latent space that represents a different diagnosis. It then generates an X-ray from that new point. This allows a doctor to see a "what if" scenario: "This is your patient's X-ray, and this is what it would look like if their heart size were normal."

Datasets

The study utilized the Indiana University hospital network dataset (Open-I), which contains 3,955 radiology reports and 7,470 chest X-rays.

  • To ensure the validity of the justification technique, the authors focused on two clear categories: "Normal" and "Cardiomegaly" (enlarged heart).
  • The final preprocessed dataset consisted of 2,446 X-rays and 4,725 textual descriptions.

Key Insights and Ideas

  • Smooth Latent Spaces are Key: By using an ARAE, the model avoids "jumps" in logic. Because the space is smooth, the model can "walk" from a diseased state to a healthy state, allowing for the generation of the nearest alternative diagnosis.
  • Counterfactuals > Heatmaps: The "aha moment" here is that humans often understand things better through comparison. Seeing the difference between the actual image and a generated "alternative" version is more intuitive for a clinician than looking at a red blob on a heatmap.
  • Multi-task Stability: Interestingly, training the model to perform multiple tasks (diagnosis, captioning, and image reconstruction) simultaneously did not hurt performance. In fact, the cycle-consistency losses helped stabilize the diagnosis loss over long training periods.

Key Results

  • Diagnostic Accuracy: The model achieved roughly 91% accuracy on the test set, performing on par with state-of-the-art single-task models like ResNet-18.
  • Captioning Quality: The model outperformed baseline convolutional captioning models across several metrics, including BLEU, ROUGE, and CIDEr.
  • Expert Validation: Three trained clinicians evaluated the system. In a blind study, they found that the proposed justification method significantly outperformed saliency maps in explaining why a diagnosis was made (scoring 2.39/4 vs. 1.31/4 for saliency maps).
  • Understanding: Clinicians felt the network had a significantly better "understanding" of the important parts of the X-ray when using the new justification method.

Limitations and Future Work

  • Resolution: Due to GPU constraints, the generated images were limited to 256x256 pixels. While sufficient for identifying an enlarged heart, this resolution might be too low for subtle pathologies like small nodules or early-stage pneumonia.
  • Class Complexity: The study focused on a binary choice (Normal vs. Cardiomegaly). Real-world radiology often involves multiple, overlapping findings. The authors note that while the method can scale to more labels, the "nearest alternative" logic becomes more complex with multiple diseases.
  • Dataset Size: The Indiana University dataset is relatively small for deep learning. Testing on larger datasets like CheXpert or MIMIC-CXR would be a logical next step.

Summary updated

Abstract

An integrated approach is proposed across visual and textual data to both determine and justify a medical diagnosis by a neural network. As deep learning techniques improve, interest grows to apply them in medical applications. To enable a transition to workflows in a medical context that are aided by machine learning, the need exists for such algorithms to help justify the obtained outcome so human clinicians can judge their validity. In this work, deep learning methods are used to map a frontal X-Ray image to a continuous textual representation. This textual representation is decoded into a diagnosis and the associated textual justification that will help a clinician evaluate the outcome. Additionally, more explanatory data is provided for the diagnosis by generating a realistic X-Ray that belongs to the nearest alternative diagnosis. With a clinical expert opinion study on a subset of the X-Ray data set from the Indiana University hospital network, we demonstrate that our justification mechanism significantly outperforms existing methods that use saliency maps. While performing multi-task training with multiple loss functions, our method achieves excellent diagnosis accuracy and captioning quality when compared to current state-of-the-art single-task methods.
BibTeX
@article{spinks2019justifying,
  author = {Graham Spinks and Marie-Francine Moens},
  title = {{Justifying diagnosis decisions by deep neural networks}},
  journal = {J. Biomed. Informatics},
  year = {2019},
  url = {https://openreview.net/forum?id=V_vChQ_MmKG},
}

Loading related content...

Figure