Raqamli Transformatsiya va Sun’iy Intellekt
AUTOMATED DIAGNOSIS OF BREAST CANCER FROM MRT IMAGES USING DEEP LEARNING ALGORITHMA
1 Tashkent University of Information Technologies named after Muhammad al-Khwarizmi.
2 Department of General Professional Sciences of the Mamun University.
Abstract
Breast cancer remains one of the most prevalent oncological diseases globally, and its early diagnosis is critical for improving survival rates. This study investigates the efficacy of deep learning algorithms for the automated diagnosis of breast cancer using magnetic resonance imaging (MRI). A dataset of 400 MRI images was utilized, comprising 250 images with cancerous regions and 150 normal images, with 2800 cancerous regions annotated. The images underwent preprocessing with Reinhard color normalization to address brightness variations. The dataset was split into 70% training, 20% validation, and 10% testing sets. A deep learning model was designed, featuring seven convolutional blocks, each with 2D convolution, ReLU activation, MaxPooling, dropout, and batch normalization, followed by feature concatenation. A recurrent layer, using LSTM to capture temporal dependencies, processed the concatenated features, which were then passed through dense layers, dropout, batch normalization, and a softmax output for classification. Two advanced object detection models, YOLOv8 and RetinaNet-DDOD, were tested. YOLOv8 achieved a mean Average Precision (mAP@0.5) of 87.5% with faster processing (0.05 seconds), while RetinaNet-DDOD outperformed it with 92.8% mAP@0.5, demonstrating higher accuracy. Mathematical models, including CloU and focal loss functions, enhanced detection of cancerous regions. RetinaNet-DDOD excelled in identifying dense and small lesions, whereas YOLOv8 performed better on simpler images. Healthy images were detected with 93.3% accuracy, and cancerous images with 88.0%, though an 8.0% false negative rate highlights challenges with small lesions. Compared to traditional manual analysis, automated systems showed superior speed and precision, underscoring their potential in clinical diagnostics and suggesting future improvements through dataset expansion.
References
No references listed for this article.