An enhanced deep learning framework for prostate cancer detection using modified VGG16 and LeNet-MobileNetV2 integration
T J Nandhini
Abstract
• An effective optimal and hybrid Deep learning model using LeNet and MobileNet v2 algorithm for the prediction purpose of prostate cancer from MRI image. • Initially, input MRI image data was taken and preprocessing step is carried for removing resilient and unwanted noise from image thereby making them augmented to use further. After that, the segmentation process was achieved using E-ACM at which the areas affected were segmented. • An extraction of features takes place using Deep-transfer Vgg16 after which the optimal selection of feature was carried. The hyper-parameter tuning was attained by employing optimization model. For this, SM-FLOA model was employed. The hyper-parameters are thus tuned and best fitness function will be attained by employing this which also enhances the rate of classifier performance. • The Hybrid Dense Auxiliary MobileNet V2 & LeNet classifier was employed for predicting & classifying prostate cancer from MRI images. • Performance assessment was made in terms of various metrics like precision; accuracy; F1-measure; recall; Error rates like MAE, RMSE, RAE, and RRSE; Kappa statistics; PPV; score; LFPR; Dice & VD. Prostate cancer remains a significant global health challenge for men, underscoring the critical need for early detection. This paper presents a novel detection framework based on a customized neural network that integrates concepts from VGG16 and a hybrid LeNet-MobileNetV2 architecture. Improvements to VGG16 aim to reduce complexity and stop the model from overfitting and joining it with LeNet and MobileNetV2 helps by having both layers focus on important features from different lightweight convolution methods. The steps for optimizing model hyperparameters were carried out in a way that could be repeated and depended for accuracy. The proposed method was trained using the annotated MRI data and then externally validated on the PROMISE12 dataset to check if it could work for different cases. For accuracy, it performed at 96.3%, for precision it reached 95.8%, for Recall: 96.7%, F1-score: 96.2% and outperformed ResNet50 and EfficientNetB0. These outcomes demonstrate that AI can produce better and quicker results in recognizing prostate cancer than is possible with present deep learning tools.