Abstract:

Recent evolution in deep learning has proven its value for CT-based lung nodule classification. Most current techniques are intrinsically black-box systems, suffering from two generalizability issues in clinical practice. First, benign-malignant discrimination is often assessed by human observers without pathologic diagnoses at the nodule level. We termed these data as "unsure-annotation data". Second, a classifier does not necessarily acquire reliable nodule features for stable learning and robust prediction with patch-level labels during learning. In this study, we construct a sure-annotation dataset with pathologically-confirmed labels and propose a collaborative learning framework to facilitate sure nodule classification by integrating unsure-annotation data knowledge through nodule segmentation and malignancy score regression. A loss function is designed to learn reliable features by introducing interpretability constraints regulated with nodule segmentation maps. Furthermore, based on model inference results that reflect the understanding from both machine and experts, we explore a new nodule analysis method for similar historical nodule retrieval and interpretable diagnosis. Detailed experimental results demonstrate that our approach is beneficial for achieving improved performance coupled with trustworthy model reasoning for lung cancer prediction with limited data. Extensive cross -evaluation results further illustrate the effect of unsure-annotation data for deep-learning based methods in lung nodule classification.

Fig. 1. Some examples of model interpretation. The first row shows different nodule inputs from unsure-annotation data and sure-annotation data, where the yellow contours on unsure-annotation data inputs are nodule segmentation by radiologists. The second row uses the CAM (Colormap Jet) to interpret which parts of the input contribute to the final prediction during ResNet reasoning. The third row shows the CAMs of our proposed model. Both the input images and CAMs are taken from the middle cross-sections of their 3D patches. The “Label” of unsure-annotation data is manually assigned by a malignancy score threshold 3 here. “GT” represents the ground truth of sure-annotation data that is confirmed by pathologic examination. For cases (a) and (e), although ResNet predicts “benign” on benign nodules correctly, this correct prediction comes from the misinterpreted evidence apart from the nodule regions. For cases (b) and (f), ResNet misclassifies benign nodules to “malignant”, where the CAM highlights both nodule and background regions. Our model can activate reliable features that are faithful to the nodule regions.

Fig. 2. Comparison of CAMs generated by models optimized with CSL or ad-CSL using different margin parameters. δ denotes the margin parameter with a bias toward nodule regions while δ′ represents that with a bias toward background regions. CAMs in the left and right columns of each nodule image come from the model using CSL and ad-CSL, respectively.

Fig. 3. Sample images of LIDC-IDRI (unsure-annotation data) nodules. Nodules with different average malignant scores are randomly selected after pre-processing and illustrated here.

DOI: 10.1016/j.media.2022.102627