Author: Hui-Shan Jian, Yu-Ying Lin 👨🔬
Affiliation: Department of Medical Imaging and Intervention, Chang Gung Memorial Hospital at Linkou 🌍
Purpose: The image quality assurance of mammographic images is crucial for correct diagnosis. To develop and validate an explainable deep-learning classifier for phantom image quality assessment of digital breast tomosynthesis (DBT). Further, it reduces the human workload and the difference of objective factors.
Methods: A total of fifty American College of Radiology digital mammography phantom images were acquired using a Hologic mammography system, and two medical physicists assessed the image quality scores, including visible, semi-invisible, and invisible. The proposed classifier was a long-short-term memory (LSTM) based framework in which the input dataset, including each phantom image of DBT, was cropped and divided into three groups: masses, specks, and fibers. The images from these three groups were passed into the encoder, and the output sequence of embeddings was fed into an LSTM layer, utilizing attention to pool these embeddings into a single feature representation. Finally, the classifier provided a semantical explanation about the possibility of classification.
Results: No significant differences were found in the predicted scores of images from each group between the expert observations and the proposed classifier. This indicates that the performance of the classifier aligns with the average scores given by the experts. The inference for the explainable classifier presented the predictions with an additional semantical explanation, which could potentially be visible, semi-invisible, or invisible.
Conclusion: The explainable classifier may improve the consistency and reproducibility of DBT image quality assurance while lessening the workload for medical physicists.