LWFDTL: lightweight fusion deep transfer learning for oral Squamous cell Carcinoma diagnosis using Histopathological oral Mucosa

Oral squamous cell carcinoma (OSCC) is among the most lethal types of cancers, especially realizing the significance of timely diagnosis. Insufficient training and awareness among primary healthcare practitioners might result in more difficult treatments, longer hospital stays, and worse survival rates. This study aims to develop a lightweight model that diagnoses cancer in its early stages and incorporates deep transfer training methods with strategies based on convolution networks and domain-specific features of OSCC samples. This work proposes a new light sequential fusion model based on histopathological data for the diagnosis of OSCC with oral mucosa images. To boost detection performance, EfficientNet is incorporated with convolutional neural networks (CNNs) as well as with max-pooling layers. The application of max-pooling also brings down the spatial features and takes care of edges and texture which proves to be both curtailing the computational workload and enhancing the performance of the model. The legitimacy of the developed framework was assessed on 1224 histopathological images where images were categorized under normal and OSCC classes depending on malignancy grades. Moreover, to improve the model training, the framework applies fine preprocessing and augmentation of input parameters such as rotations, zooms, flips, contrasts, and crops. These augmentations help the model learn from different views and hence boost its detection capability across all the aspects of the data. Experimental findings shows that the proposed framework had an accuracy before enhancement of 91.0% and an accuracy after enhancement of 98.9%, which is even higher than such methods as the traditional single-model and the majority of modern methods. Additionally, our model is lightweight, requiring significantly less time than 0.05 segments on an average GPU. It is also conclude that the proposed model be run in clinical settings with modest computational capabilities. These results demonstrate the effectiveness of combined methods in revealing an innovative, efficient, and affordable diagnostic apparatus for OSCC that will help in detecting the illness at an early stage and aid in the development of a proper treatment plan.