Oral Presentation at the European Congress of Radiology, Vienna, 2019

Purpose

To explain predictions of a deep residual convolutional network for characterization of lung nodule by analyzing heat
maps

Methods and Materials

A 20-layer deep residual CNN was trained on 1245 Chest CTs from NLST trial to predict the malignancy risk of a nodule. We used occlusion to systematically block regions of a nodule and map drops in malignancy risk score to generate clinical attribution heatmaps on 160 nodules from LIDC-IDRI dataset, which were analysed by a thoracic radiologist. The features were described as heat inside nodule (IH)-bright areas inside nodule, peripheral heat (PH)-continuous/interrupted bright areas along nodule contours, heat in adjacent plane(AH)-brightness in scan planes juxtaposed with the nodule, satellite heat (SH)- a smaller bright spot in proximity to nodule in the same scan plane, heat map larger than nodule (LH)-bright areas corresponding to the shape of the nodule seen outside the nodule margins and heat in calcification (CH)

Results

These six features were assigned binary values. This feature vector was fed into a standard J48 decision tree with 10-fold cross-validation, which gave an 85 % weighted classification accuracy with a 77.8 %TP rate, 8% FP rate for benign cases and 91.8% TP and 22.2 %FP rates for malignant cases. IH was more frequently observed in nodules classified as malignant whereas PH, AH, and SH were more commonly seen in nodules classified as benign.

Conclusion

We discuss the potential ability of a radiologist to visually parse the deep learning algorithm generated 'heat map' to identify features aiding classification