Background and SOTA Results
Here we highlight a few state-of-the-art results for breast ultrasound classification.
Classification models are typically trained in two different ways: - Strongly-supervised approach each individual image is labeled as malignant or benign.
- Weakly-supervised approach each exam, consisting of multiple images, is labeled as malignant or benign.
The strongly-supervised approach requires time consuming and expensive annotation by radiologists who must review each individual image after acquiring biopsy results. The weakly-supervised approach requires only that a breast exam has corresponding biopsy results or that a sufficient amount of time has passed so that the breast may be assume to have no malignancy.
State-of-the-art results for strongly-supervised classification.
In a 2023 comparative study Ferreira et al. (2023) combined two small public datasets (810 total images) to show that EfficientNet can be used to attain Accuracy = 92.59% and AUC = 93.53% on their validation set.
In the 2022 article “Deep learning based on ultrasound images assists breast lesion diagnosis in China: a multicenter diagnostic study” Gu et al. (2022) collected and annotated 14,043 ultrasound images and trained a VGG19 model to achieve, on a separate test set, Accuracy = 86.4%, AUC = 91.3%, Sensitivity = 88.84%, and Specificity = 83.77%.
State-of-the-art results for weakly-supervised collection.
In the 2021 article “Artificial intelligence system reduces false-positive findings in the interpretation of breast ultrasound exams,” Shen et al. (2021) collected 288,767 exams consisting of 5,442,907 images (about 10% of the exams were associated with biopsies). They trained a multiple instance (weakly-supervised) model using a Resnet18 backbone and only exam-level labels to achieve an AUC of 97.6% on their test set (AUC = 94.0% on biopsied population). Moreover, at 90.1% sensitivity, their model demonstrated 85.6% specificity.