Colorectal cancer (CRC), also known as colon cancer, is a significant global health concern, ranking as the third most common cause of cancer death. Early detection is crucial for improving patient outcomes. While traditional methods like endoscopy exist, they can be inaccurate and uncomfortable. Computer-aided diagnosis (CAD) systems offer a promising alternative. Recent advances in machine learning, particularly deep learning (DL), have shown potential for improving the accuracy and efficiency of CRC detection. This paper focuses on developing an optimized DL model for colon cancer classification and segmentation, addressing limitations in existing research such as limited datasets and the lack of optimization algorithm incorporation in most CNN-based approaches. The study aims to improve the accuracy of colon cancer detection and classification by using optimized deep learning approaches, particularly convolutional neural networks (CNNs) with different optimizers, and by making a real colonoscopy dataset available as an open-source resource. The main objective is to explore the potential of various deep learning techniques for improving the accuracy and efficiency of colon cancer classification.

The paper reviews existing literature on colon cancer detection and classification, highlighting the shift from traditional machine learning methods (SVM, K-nearest Neighbor, Random Forest) with handcrafted feature extraction to deep learning approaches. Various imaging techniques (histopathology, endoscopy, colonoscopy) and deep learning architectures (CNNs, including VGG16, ResNet, Inception, DenseNet) are discussed. The review emphasizes the limitations of existing studies, such as reliance on CNNs without optimization algorithms, small and low-quality datasets, and the lack of publicly available datasets for certain imaging techniques. Specific prior studies on using CNNs for CRC detection and classification with various pre-trained models (e.g., VGG16, ResNet) and the utilization of techniques like transfer learning and data augmentation are reviewed. The authors highlight the contribution of their research to address the limitations in existing work by systematically evaluating the performance of different deep learning optimizers and by providing a new dataset.

The study utilizes convolutional neural networks (CNNs) as the core deep learning architecture for colon cancer classification. Six state-of-the-art optimization algorithms are investigated: Stochastic Gradient Descent (SGD), Adamax, AdaDelta, Root Mean Square Propagation (RMSprop), Adaptive Moment Estimation (Adam), and Nadam (Nesterov-accelerated Adam). Each optimizer is paired with a CNN model and trained on four datasets: Colon, Colonoscopy, Warwick-QU, and CRC-VAL-HE-7K. Detailed descriptions of each optimizer are provided, including mathematical formulas for weight updates. Image preprocessing steps including resizing (224x224x3), normalization, and augmentation are employed. Data augmentation techniques such as rotation, scaling, flipping, and cropping are used to increase the size of the training datasets, especially for those with a limited number of original images. The performance of each CNN-optimizer combination is evaluated using several metrics: precision, recall, F1-score, and accuracy. The experiments are conducted using Python with Keras and TensorFlow on Google Colab, utilizing GPUs for faster training. Training and validation curves (loss and accuracy) are generated and analyzed to identify overfitting or underfitting issues.

The CNN-Adam optimizer consistently demonstrated superior performance across the four datasets, achieving an average accuracy of 82%. Dataset 1 (Colon) yielded the highest accuracy results, with CNN-Adam achieving 95%, CNN-RMSprop 76%, and CNN-Adadelta 96%. Other datasets exhibited varied performance across different optimizers. The Colonoscopy dataset showed the best results with CNN-Nadam (79% accuracy). The Warwick-QU dataset showed no single optimizer dominating, with CNN-Adam (F1-score 0.72), CNN-SGD (Recall 0.78), and CNN-Adamax (Precision 0.77) each excelling in different metrics. The CRC-VAL-HE-7K dataset showed CNN-Adam as the best performer (90% accuracy). The analysis of training and validation curves provided insights into the learning process for different optimizers, revealing the effectiveness of Adam and Nadam in minimizing loss and maximizing accuracy. Data augmentation had a varied effect on the performance, suggesting that a careful balance is needed between augmentation and data quality.

The results demonstrate the effectiveness of deep learning, specifically CNNs with the Adam optimizer, for classifying colon cancer images. The superior performance of Adam compared to other optimizers suggests its suitability for this specific task. The variation in performance across different datasets highlights the importance of dataset characteristics and the need for robust models. The finding that data augmentation did not always improve performance suggests that careful consideration should be given to the augmentation strategy employed, and perhaps a better strategy is required to improve the results for smaller datasets. The study's findings contribute to the development of efficient and accurate CAD systems for CRC, facilitating early detection and potentially improving patient outcomes. The provision of the real colonoscopy dataset as an open-source resource will benefit the research community.

This research successfully developed and evaluated optimized deep learning models for colon cancer detection and segmentation. The CNN-Adam model consistently outperformed other optimizers across multiple datasets, demonstrating its potential for accurate and efficient colon cancer classification. Future work should focus on expanding the dataset, including more diverse images and clinical information, to improve model generalizability and reliability. Further investigation into other deep learning architectures and optimization techniques could also enhance performance. External validation by oncologists and clinical trials are essential to translate the findings into clinical practice.

The study's findings are limited by the specific datasets used, and the generalizability to other datasets or clinical settings needs further validation. The number of images in some datasets was relatively small, even after augmentation, which could impact the model's generalization ability. The study's focus was on image classification; future research should also investigate the efficacy of the system for polyp detection and segmentation. Finally, the study did not consider other important aspects such as the real-time performance needed for clinical implementation.