Research On Deep Learning-assisted Diagnosis Methods For Medical Image

The artificial intelligence(AI)technology,represented by deep learning,has gained popularity in the research community and has been utilized throughout the entire process of medical image processing and analysis.Computer-aided diagnosis based on deep learning methods effectively supports physicians in making decisions,leading to significant improvements in efficiency and accuracy.This paper takes cervical cancer precancerous lesions and ovarian cysts as two different disease diagnoses as the starting point,and focuses on the enhancement of cervical cancer precancerous lesion images,ovarian cyst image classification,and lesion segmentation of both diseases.The research contents are as follows:To address the issue of low image quality of cervical cancer precancerous lesions under complex interference,an image enhancement framework based on conditional entropy generative adversarial networks is proposed.A Retinex network with decomposition function is established to obtain the reflection image of low-quality images,and a conditional generative adversarial network is designed to enhance the quality of the reflection image.A loss function on conditional entropy distance is constructed to alleviate the overfitting phenomenon during the training process.Experimental results show that the structural similarity coefficient and peak signal-to-noise ratio indicators of the proposed method are better than other algorithms,and it can significantly enhance image quality while preserving image details.To address the issue of rapid intelligent diagnosis of ovarian cyst ultrasound images,a lightweight deep learning classification model(Ocys-Net)is proposed,which can classify normal pelvic areas,ovarian cysts,and non-pure ovarian cysts.The model adopts a reverse bottleneck design strategy to enhance feature extraction capabilities and uses an efficient channel attention(ECA)module to focus on the features of pathological information,achieving local cross-channel interaction and mitigating the defects brought by channel dimensionality reduction,which improves learning efficiency.Experimental results show that the proposed method achieves an accuracy of 95.93% on the ovarian cyst dataset,demonstrating its practical clinical utility.To address the issue of accurate lesion segmentation in cervical cancer precancerous lesions and ovarian cyst images,an image segmentation algorithm based on the Swin-Unet model is proposed.The Swin-Unet network model is used for preliminary prediction of medical image lesions,and an improved conditional random field(CRF)model is combined to correct the prediction results to ensure segmentation continuity and accuracy.Furthermore,the interpretability of the proposed method is explored by using the gradient-weighted class activation mapping(Grad-CAM)algorithm for segmentation model visualization,which provides key explanations for the feature attention and attention range of the segmentation model towards different types of data.Experimental results show that the proposed method achieves fine lesion segmentation in cervical cancer precancerous lesions and ovarian cyst images,to some extent with interpretability,effectively improving the reliability,accuracy,and trustworthiness of the segmentation model.