A Hybrid Lightweight Deep Learning Framework for Colorectal Cancer Detection Using an Adaptive Gated Feature Fusion and Grid Search-Based Hyperparameter Optimization.
Authors
Affiliations (6)
Affiliations (6)
- Department of Software Engineering, Nile University of Abuja, Abuja, Nigeria.
- Department of Computer Science, Federal University, Oye Ekiti, Ekiti State, Nigeria.
- Department of Multimedia Engineering, Kaunas University of Technology, Kaunas, Lithuania. [email protected].
- Department of Computer Systems Engineering, Tshwane University of Technology, Pretoria, South Africa. [email protected].
- Department of Computer Science, Redeemer's University, Osun State, Ede, Nigeria. [email protected].
- Department of Computer Systems Engineering, Tshwane University of Technology, Pretoria, South Africa.
Abstract
The timely and precise diagnosis of colorectal cancer is a critical area for enhancing the quality of care for patients, which could lead to better outcomes. Despite the fact that deep learning has demonstrated great potential in the field of histopathological image analysis, the accuracy of models can be influenced by various factors, including network architecture, optimization method, and data imbalance. The aim of this study is to explore the impact of various optimization algorithms on the classification of colorectal cancer in CRC-VAL-He-7 K database. A hybrid lightweight (HYL) model was designed and tested using the CNN, DeeperCNN, ResNetLike, DenseNet121, EfficientNetB0, and L2RMLP architectures. For the purpose of improving model reliability, it was decided to use grid search for the automatic tuning of the hyperparameters, and class imbalance issues were tackled by the following data augmentation and class-weighted learning techniques: five optimizers: Adam, AdamW, RMSprop, SGD, and Adadelta were evaluated in identical experiments. Both model architecture and the choice of optimizer had a profound influence on the results of classification. The proposed hybrid lightweight model provided competitive accuracy while requiring fewer computational resources, whereas EfficientNetB0 and DenseNet121 exhibited the best overall performance. The results of statistical analysis also showed significant differences between the model-optimizer combinations evaluated. The results offer real-world applications for designing effective and accurate deep learning systems for colorectal cancer diagnosis and highlight the need for optimization and data-balancing approaches in medical image classification.