Interpretative Computer-aided Lung Cancer Diagnosis: from Radiology Analysis to Malignancy Evaluation

Shaohua Zheng, Zhiqiang Shen, Chenhao Pei, Wangbin Ding, Haojin Lin, Jiepeng Zheng, Lin Pan, Bin Zheng, Liqin HuangarXiv 2021 · DOI · OpenReview

TL;DR

This paper introduces R2MNet, a novel deep learning architecture that combines nodule radiology analysis (classification of nodule types) with malignancy evaluation for lung cancer diagnosis. By extracting radiological features and using them to guide malignancy assessment, R2MNet achieves higher accuracy and provides interpretable explanations, mimicking the diagnostic process of experienced radiologists.

What Problem Does This Paper Address?

Lung cancer is a leading cause of cancer death, and low-dose computed tomography (LDCT) screening is effective in reducing mortality. However, manually analyzing CT scans is time-consuming and relies heavily on radiologist experience. While deep learning has shown promise in nodule detection, current methods for evaluating nodule malignancy often lack interpretability, providing only a probability score without explaining why a nodule is deemed malignant. This "black box" nature hinders clinical adoption. This paper addresses the need for a more interpretable and clinically aligned approach to lung nodule malignancy evaluation by integrating radiological feature analysis into the deep learning pipeline.

Methods

The core of this work is the R2MNet architecture, which comprises two main components:

  • RNet (Radiology Analysis Network): This network classifies pulmonary nodules into four categories: solid nodules (SN), ground-glass opacity nodules (GGO), mixed GGO nodules (MGGO), and calcified nodules (CN). This step aims to extract clinically relevant radiological features.
  • MNet (Malignancy Evaluation Network): This network evaluates the likelihood of a nodule being malignant.

A key innovation is the Assisted Gating Unit (AGU), which acts as a bridge between RNet and MNet. AGUs transform the radiological features extracted by RNet into a "channel descriptor." This descriptor is then used to guide MNet, helping it focus on the most informative regions and features for malignancy assessment. This is inspired by attention mechanisms and aims to mimic how radiologists use specific visual cues.

For interpretability, the authors propose Channel-Dependent Activation Mapping (CDAM). CDAM visualizes which parts of the nodule the network is focusing on by analyzing the dependencies between activation maps. This provides insights into the model's decision-making process, highlighting features like shape and density that are crucial for malignancy assessment, aligning with human expert cognition.

Datasets

The study utilizes the LIDC-IDRI (Lung Image Database Consortium and Image Database Resource Initiative) dataset. This dataset is a widely used benchmark for lung nodule analysis. The authors selected a subset of 1004 nodules from the LUNA16 challenge, with nodules labeled as malignant if their average malignancy score was >3 and benign if <3. Some nodules were excluded due to ambiguous scores or limited radiologist annotations.

Key Insights and Ideas

  • Radiology Analysis as a Precursor: Integrating nodule classification (radiology analysis) as an intermediate step before malignancy evaluation significantly improves performance and interpretability. It provides a structured way to extract clinically relevant features.
  • AGU for Feature Guidance: The AGU module effectively channels the information from the radiology analysis network to the malignancy evaluation network, acting as a learned attention mechanism that focuses the malignancy assessment on relevant nodule characteristics.
  • CDAM for Clinical Alignment: The proposed CDAM method provides visual explanations that not only highlight important nodule regions but also demonstrate that the model's focus (e.g., on shape and density) aligns with how experienced radiologists diagnose malignancy. This builds trust and understanding.
  • Multi-task Learning Synergy: The joint training of RNet and MNet in a multi-task learning framework allows the tasks to reinforce each other, leading to better overall performance than training either task in isolation.

Key Results

  • Radiology Analysis Performance: R2MNet achieved an Area Under the Curve (AUC) of 96.27% for nodule radiology analysis, outperforming the standalone RNet.
  • Malignancy Evaluation Performance: R2MNet achieved an AUC of 97.52% for nodule malignancy evaluation, surpassing the standalone MNet.
  • Interpretability: CDAM visualizations confirmed that R2MNet focuses on nodule shape and density, which are known critical factors for malignancy assessment by radiologists.
  • Stability: The R2MNet model demonstrated more stable performance across cross-validation compared to other models, suggesting robustness.

Limitations and Future Work

  • Dependency on Nodule Detection: The performance of R2MNet is contingent on the accuracy of the initial nodule detection step. The paper acknowledges that the input data is derived from either manual selection or automatic nodule detectors.
  • Future Work: The authors suggest developing an end-to-end CAD system that integrates nodule detection, segmentation, and malignancy prediction. This would create a more comprehensive and clinically applicable system.

Summary updated

Abstract

Background and Objective:Computer-aided diagnosis (CAD) systems promote diagnosis effectiveness and alleviate pressure of radiologists. A CAD system for lung cancer diagnosis includes nodule candidate detection and nodule malignancy evaluation. Recently, deep learning-based pulmonary nodule detection has reached satisfactory performance ready for clinical application. However, deep learning-based nodule malignancy evaluation depends on heuristic inference from low-dose computed tomography volume to malignant probability, which lacks clinical cognition. Methods:In this paper, we propose a joint radiology analysis and malignancy evaluation network (R2MNet) to evaluate the pulmonary nodule malignancy via radiology characteristics analysis. Radiological features are extracted as channel descriptor to highlight specific regions of the input volume that are critical for nodule malignancy evaluation. In addition, for model explanations, we propose channel-dependent activation mapping to visualize the features and shed light on the decision process of deep neural network. Results:Experimental results on the LIDC-IDRI dataset demonstrate that the proposed method achieved area under curve of 96.27% on nodule radiology analysis and AUC of 97.52% on nodule malignancy evaluation. In addition, explanations of CDAM features proved that the shape and density of nodule regions were two critical factors that influence a nodule to be inferred as malignant, which conforms with the diagnosis cognition of experienced radiologists. Conclusion:Incorporating radiology analysis with nodule malignant evaluation, the network inference process conforms to the diagnostic procedure of radiologists and increases the confidence of evaluation results. Besides, model interpretation with CDAM features shed light on the regions which DNNs focus on when they estimate nodule malignancy probabilities.
BibTeX
@misc{zheng2021interpretative,
  author = {Shaohua Zheng and Zhiqiang Shen and Chenhao Pei and Wangbin Ding and Haojin Lin and Jiepeng Zheng and Lin Pan and Bin Zheng and Liqin Huang},
  title = {{Interpretative Computer-aided Lung Cancer Diagnosis: from Radiology Analysis to Malignancy Evaluation}},
  year = {2021},
  eprint = {2102.10919},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.10919},
}
Figure