E1487. Investigation of the Usefulness of Vision Transformer: An Artificial Intelligence Model for Mammography Diagnosis
Authors
Akihito Nakajima;
Saitama Medical University International Medical Center
Yoko Usami;
Saitama Medical University International Medical Center
Hiroyuki Tajima;
Saitama Medical University International Medical Center
Kazuo Matsuura;
Saitama Medical University International Medical Center
Akihiko Osaki;
Saitama Medical University International Medical Center
Objective:
To evaluate the usefulness of the Vision Transformer (ViT), a transformer-based model, in mammography using ViT. We examined the usefulness of ViT in mammography.
Materials and Methods:
The DDSM (Digital Database for Screening Mammography: The DDSM (Digital Database for Screening Mammography: www.kaggle.com/skooch/ddsm-mammography) consists of 55,890 crops of training data, 14% of which are prevalent cases and 86% of which are normal cases. The evaluation items were (1) label normal: 0 negative /1 positive (2) label: 0 negative, 1 benign calcification, 2 benign mass, 3 malignant calcification, 4 malignant mass. The DDSM was used to create a learning model by transition learning using ViT. After creating the model, we tested it on our own data (192 images, 48 cases). In addition, two readers (Reader A and Reader B) were tested to compare the results.
Results:
ViT was 0.83/0.83/0.83/0.83 in accuracy/precision/recall/F1/Cohen Kappa for good and bad classification, and 0.83/0.83/0.83/0.83 in F1/Cohen Kappa for good and bad classification.0.83/0.83/0.83/0.83/0.83/0.57, Reader A was 0.93/0.93/0.93/0.93/0.93/0.82, and Reader B was 0.88/0.88/0.88/0.88/0.87/0.66. In the five categories, ViT was 0.49/0.61/0.49/0.54/0.17, Reader A was 0.89/0.91/0.89/0.89/0.89/0.79, and Reader B was 0.77/0.79/0.77/0.77/0.77/0.55, respectively.
Conclusion:
ViT performed as well as the readers in the benign/malignant classification, but could not surpass them in the categorical classification.
