Explainability Through Human-Centric Design for XAI in Lung Cancer Detection

Amy Rafferty, Rishi Ramaesh, Ajitha RajanIJCAI 2025 · DOI

TL;DR

This paper introduces XpertXAI, a novel concept bottleneck model for lung cancer detection that leverages expert-defined clinical concepts to improve explainability and accuracy. Unlike existing methods that often produce clinically irrelevant or misleading explanations, XpertXAI aligns better with expert reasoning, demonstrating the value of human-centric design in medical AI.

What Problem Does This Paper Address?

Deep learning models have shown great promise in medical imaging, particularly for detecting diseases like lung cancer from chest X-rays. However, their "black-box" nature makes it difficult for clinicians to trust and understand their decisions, hindering widespread adoption. Existing explainable AI (XAI) techniques, both post-hoc (applied after training) and ante-hoc (built-in interpretability), often fall short in medical contexts. Post-hoc methods like LIME and Grad-CAM can highlight irrelevant regions or miss crucial features, while even advanced multimodal models struggle to capture pathology-specific signals recognized by experts. The authors hypothesize that a critical limitation is the lack of expert involvement in the design of these XAI techniques.

Methods

The core of this work is XpertXAI, an expert-driven Concept Bottleneck Model (CBM). CBMs work in two stages: first, they predict intermediate "concepts" from an image, and then these concepts are used to predict the final label. XpertXAI distinguishes itself by:

  • Expert-Defined Concepts: Instead of learning concepts unsupervised (like in XCBs), XpertXAI uses clinical concepts curated in direct collaboration with radiologists. These concepts are extracted from radiology reports using NLP techniques and are validated by experts. This ensures the concepts are clinically meaningful and directly interpretable.
  • Multi-Pathology Training: While the clinical focus is lung cancer, XpertXAI is trained on a dataset encompassing multiple thoracic pathologies. This mirrors real-world diagnostic scenarios and tests the generalizability of the approach.
  • Architecture: XpertXAI uses an InceptionV3 CNN for concept prediction and a Decision Tree (DT) classifier for label prediction, finding DTs to perform best for this task.

The authors compare XpertXAI against:

  • Post-hoc Image XAI: LIME, SHAP, and Grad-CAM applied to a standard InceptionV3 model.
  • Post-hoc LLM-based XAI: CXR-LLaVA, a multimodal large language model for chest X-rays.
  • Ante-hoc CBM: XCBs, an unsupervised CBM.

Datasets

  • MIMIC-CXR: A large-scale public dataset of chest X-rays and associated radiology reports used for both training and evaluation. The authors focus on the Posterior-Anterior (PA) view and select six clinically significant labels (Healthy, Lung Cancer, Pneumonia, Pneumothorax, Pleural Effusion, Cardiomegaly). They employ undersampling to address class imbalance.
  • VinDr-CXR: A subset of this dataset was used for evaluating the clinical relevance of image-based XAI methods by comparing explanations to radiologist-annotated bounding boxes.

Key Insights and Ideas

  • Expert Input is Crucial for Clinical XAI: The most significant takeaway is that directly involving domain experts in defining the interpretable concepts is paramount for creating XAI methods that are truly useful in clinical practice. Unsupervised or purely data-driven concept discovery often leads to clinically irrelevant outputs.
  • Concept Bottleneck Models are Promising, but Need Guidance: CBMs offer a strong framework for interpretability, but their effectiveness hinges on the quality and clinical relevance of the concepts they learn or are given.
  • Post-hoc Methods Struggle with Clinical Relevance: Image-based post-hoc methods like LIME, SHAP, and Grad-CAM consistently failed to align with expert opinions and often highlighted anatomically implausible regions, raising serious concerns about their utility in critical diagnostic tasks.
  • LLMs Can Hallucinate in Medical Contexts: CXR-LLaVA, while capable, showed vulnerability to prompt-induced hallucinations and often produced false negatives or emphasized less critical findings.
  • Human-Centric Design Scales: The success of XpertXAI in a multi-pathology setting suggests that this human-centric design principle can be extended beyond single-disease detection to broader diagnostic contexts.

Key Results

  • Concept Prediction: XpertXAI achieved a concept F1 score of 0.842, outperforming XCBs (0.799) and CXR-LLaVA (0.658), supporting the hypothesis that expert input enhances interpretability.
  • Label Prediction: XpertXAI with a Decision Tree classifier achieved the highest F1 score of 0.866 on the multiclass MIMIC-CXR test set, surpassing the InceptionV3 baseline and XCBs.
  • Expert Evaluation (Lung Cancer): A radiologist rated XpertXAI explanations as significantly more accurate, reliable, and aligned with expert reasoning compared to all other evaluated methods. XpertXAI explanations were often deemed "fully clinically relevant" (score of 3), whereas post-hoc methods predominantly received scores of 0 or 1.
  • Qualitative Examples: Figure 3 vividly illustrates XpertXAI correctly identifying "Mass" and "Irregular Hilum" for a cancerous X-ray, while other methods produced less informative or misleading explanations.

Limitations and Future Work

  • Limited Expert Evaluation: The current expert evaluation, while insightful, was performed on a small subset of cases (60 X-rays). The authors plan to expand this in future work.
  • Generalizability of Architecture: While XpertXAI uses an InceptionV3 backbone, the authors acknowledge that the generalizability of this specific architecture might be a concern, though it has performed well in the field.
  • Future Directions: The authors aim to adapt their work to other imaging modalities like CT and MRI, and to further refine the expert evaluation process.

Summary updated

BibTeX
@inproceedings{rafferty2025explainability,
  author = {Amy Rafferty and Rishi Ramaesh and Ajitha Rajan},
  title = {{Explainability Through Human-Centric Design for XAI in Lung Cancer Detection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1147},
}
Figure