Abstract
This work presents SE-PCOSNet, a novel deep learning PCOS diagnosis model driven by ultrasound image analysis. The architecture extends the fundamental Convolutional Neural Network (CNN) with Squeeze-and-Excitation (SE) blocks that support dynamic feature channel recalibration. The addition of the SE blocks supports the model’s ability to detect weak and faint features in gynecological ultrasound images, crucial to successful PCOS diagnosis. The suggested architecture consists of three convolutional blocks, each of which includes SE blocks for facilitating improved discriminative feature extraction. Apart from architectural enhancement, the technique utilizes state-of-the-art data preprocessing, data augmentation, and attention mechanisms. GPU acceleration, regularization techniques, and crossvalidation enhance the training process to render it robust and avoid over fitting. With a vast and heterogeneous dataset of over 1,900 pelvic ultrasound images, SE-PCOSNet achieved impressive external validation performance with 82.6% precision and overall, 100% recall for both PCOS-positive and PCOS-negative classes. The model also has about twice the computational speed of the standard CNN models and achieves this without sacrificing diagnostic accuracy. The results affirm the strength and efficiency of SE-PCOSNet in handling actual clinical data sets. The use of SE blocks not only allows for feature recalibration but also enables remarkable sensitivity and specificity in PCOS condition classification. The model possesses great potential for integration into automated, real-time diagnosis systems, with the intention to maximize the gynecological diagnosis in clinical settings. Additional research can further expand the framework and evaluate the use of this method across a greater range of diagnostic applications.
Keywords: Medical Image Processing, PCOS, ReLU (Rectified Linear Unit), SE Blocks (Squeeze-and-Excitation Block), SoftMax, Ultrasound Imaging.