Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Concept-based Explainable Malignancy Scoring on Pulmonary Nodules in CT Images

Rinat I. Dumaev, Sergei A. Molodyakov, Lev V. Utkin

TL;DR

This work tackles the interpretability gap in CT-based pulmonary nodule malignancy scoring by integrating a CNN-based shared feature extractor with a concept predictor and a generalized additive model (GAM) decision explainer. The approach produces a final malignancy score along with additive, per-concept contributions for both numerical and categorical radiologic attributes, enabling human-friendly explanations. Using the LIDC-IDRI dataset, the model achieves competitive attribute-scoring and malignancy performance, with shape-function analyses revealing clinically intuitive behaviors (e.g., absence of calcification and higher spiculation increase malignancy risk). The framework thus provides transparent, clinically aligned decision reasoning that can support radiologists in making more informed diagnoses.

Abstract

To increase the transparency of modern computer-aided diagnosis (CAD) systems for assessing the malignancy of lung nodules, an interpretable model based on applying the generalized additive models and the concept-based learning is proposed. The model detects a set of clinically significant attributes in addition to the final malignancy regression score and learns the association between the lung nodule attributes and a final diagnosis decision as well as their contributions into the decision. The proposed concept-based learning framework provides human-readable explanations in terms of different concepts (numerical and categorical), their values, and their contribution to the final prediction. Numerical experiments with the LIDC-IDRI dataset demonstrate that the diagnosis results obtained using the proposed model, which explicitly explores internal relationships, are in line with similar patterns observed in clinical practice. Additionally, the proposed model shows the competitive classification and the nodule attribute scoring performance, highlighting its potential for effective decision-making in the lung nodule diagnosis.

Concept-based Explainable Malignancy Scoring on Pulmonary Nodules in CT Images

TL;DR

This work tackles the interpretability gap in CT-based pulmonary nodule malignancy scoring by integrating a CNN-based shared feature extractor with a concept predictor and a generalized additive model (GAM) decision explainer. The approach produces a final malignancy score along with additive, per-concept contributions for both numerical and categorical radiologic attributes, enabling human-friendly explanations. Using the LIDC-IDRI dataset, the model achieves competitive attribute-scoring and malignancy performance, with shape-function analyses revealing clinically intuitive behaviors (e.g., absence of calcification and higher spiculation increase malignancy risk). The framework thus provides transparent, clinically aligned decision reasoning that can support radiologists in making more informed diagnoses.

Abstract

To increase the transparency of modern computer-aided diagnosis (CAD) systems for assessing the malignancy of lung nodules, an interpretable model based on applying the generalized additive models and the concept-based learning is proposed. The model detects a set of clinically significant attributes in addition to the final malignancy regression score and learns the association between the lung nodule attributes and a final diagnosis decision as well as their contributions into the decision. The proposed concept-based learning framework provides human-readable explanations in terms of different concepts (numerical and categorical), their values, and their contribution to the final prediction. Numerical experiments with the LIDC-IDRI dataset demonstrate that the diagnosis results obtained using the proposed model, which explicitly explores internal relationships, are in line with similar patterns observed in clinical practice. Additionally, the proposed model shows the competitive classification and the nodule attribute scoring performance, highlighting its potential for effective decision-making in the lung nodule diagnosis.
Paper Structure (16 sections, 4 equations, 4 figures, 3 tables)

This paper contains 16 sections, 4 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Concept-Based Models
  4. Interpretable Malignancy Scoring on Pulmonary Nodules
  5. Methods
  6. Framework Architecture
  7. Shared Feature Extractor
  8. Concepts Predictor
  9. Decision Explainer
  10. Experiments
  11. Experimental Setup
  12. Dataset
  13. Implementation details
  14. Quantitative Evaluation
  15. Qualitative Evaluation
  16. ...and 1 more sections

Figures (4)

  • Figure 1: Overview of the proposed framework. Our approach is divided into several components. (a) Deep 3D CNN-based shared feature extractor encodes input image slices of the nodule to feature embedding in a low-dimensional latent space. (b) Learning numerical (e.g. subtlety, lobulation, margin) and categorical concept features (e.g. internal structure and calcification) by raw feature embedding. (c, d) Multiple GAM subnetworks are trained on each concept input feature and weighted sums are learned over the subnetworks. The outputs corresponding to each concept are summed and a bias is added to obtain the final malignancy score. (e) Predicted numerical and categorical concept values and corresponding calculated contributions are used to explain the predicted malignancy score.
  • Figure 2: GAM architecture. Multiple sub networks are trained on each input concept and weighted sums are learned over the sub networks
  • Figure 3: Concepts contributions to the malignancy risk. These plots show the individual shape functions learned by an Decision Explainer module using GAMs for each input concept.
  • Figure 4: Predicted concepts contributions of three nodules. Examples of provided concept contributions to the predicted malignancy score by our method for three nodules from Table \ref{['tab:predictions_samples']}.