One step further into the blackbox: a pilot study of how to build more confidence around an AI-based decision system of breast nodule assessment in 2D ultrasound.

ABSTRACT

OBJECTIVES: To investigate how a DL model makes decisions in lesion classification with a newly defined region of evidence (ROE) by incorporating "explainable AI" (xAI) techniques. METHODS: A data set of 785 2D breast ultrasound images acquired from 367 females. The DenseNet-121 was used to classify whether the lesion is benign or malignant. For performance assessment, classification results are evaluated by calculating accuracy, sensitivity, specificity, and receiver operating characteristic for experiments of both coarse and fine regions of interest (ROIs). The area under the curve (AUC) was evaluated, and the true-positive, false-positive, true-negative, and false-negative results with breakdown in high, medium, and low resemblance on test sets were also reported. RESULTS: The two models with coarse and fine ROIs of ultrasound images as input achieve an AUC of 0.899 and 0.869, respectively. The accuracy, sensitivity, and specificity of the model with coarse ROIs are 88.4%, 87.9%, and 89.2%, and with fine ROIs are 86.1%, 87.9%, and 83.8%, respectively. The DL model captures ROE with high resemblance of physicians' consideration as they assess the image. CONCLUSIONS: We have demonstrated the effectiveness of using DenseNet to classify breast lesions with limited quantity of 2D grayscale ultrasound image data. We have also proposed a new ROE-based metric system that can help physicians and patients better understand how AI makes decisions in reading images, which can potentially be integrated as a part of evidence in early screening or triaging of patients undergoing breast ultrasound examinations. KEY POINTS • The two models with coarse and fine ROIs of ultrasound images as input achieve an AUC of 0.899 and 0.869, respectively. The accuracy, sensitivity, and specificity of the model with coarse ROIs are 88.4%, 87.9%, and 89.2%, and with fine ROIs are 86.1%, 87.9%, and 83.8%, respectively. • The first model with coarse ROIs is slightly better than the second model with fine ROIs according to these evaluation metrics. • The results from coarse ROI and fine ROI are consistent and the peripheral tissue is also an impact factor in breast lesion classification.