The observational diagnostic study conducted at Medipol University’s clinical research ethics committee involved analyzing electronic records from March 2022 to January 2023. The study focused on asymptomatic females aged 40 years and above who underwent breast tissue screening via Digital Breast Tomosynthesis (DBT) at Medipol Mega University Hospital. A total of 2521 DBT examinations were retrospectively analyzed to select cases falling under specific imaging interpretation categories.
The inclusion criteria comprised cases classified as BIRADS 1, 2, and 3, along with BIRADS 0 cases showing BIRADS 1, 2, or 3 findings during ultrasound examinations conducted simultaneously. On the other hand, exclusion criteria involved images lacking adequate pectoral muscle visibility, suspicions of malignant masses (BIRADS 4, 5, 6), patients with cancer history or ongoing treatment, patients with diabetes prescriptions, and right medial-lateral oblique images due to expected muscle hypertrophy.
Pectoral muscle images from eligible female patients were collated in DICOM format alongside HbA1c% values obtained from same-day blood samples. The patients’ HbA1c levels categorized them into normal, prediabetic, or diabetic groups. A minimum of 1000 images per category within three age ranges (40–49, 50–59, 60+) were aimed for analysis, totaling 11,594 images post filtration.
The dataset underwent data augmentation to enhance diversity through techniques like horizontal and vertical flips and rotation. Convolutional Neural Networks (CNN), specifically EfficientNetB5 architecture, were used for image classification tasks. The model’s training involved parameter adjustments based on architecture characteristics and best practices, leveraging Google Colab’s resources for efficient learning outcomes. Performance metrics such as precision, recall, and F1-score were utilized to evaluate the model’s accuracy in distinguishing between normal, prediabetic, and diabetic statuses.