Automated System for Colon Cancer Detection and Segmentation Based on Deep Learning Techniques

The study addresses the need for early and accurate detection of colorectal cancer (CRC), a leading cause of cancer morbidity and mortality worldwide. Traditional diagnostic procedures such as endoscopy and wireless capsule endoscopy are effective but can be uncomfortable, time-consuming, and prone to human error. Computer-aided diagnosis (CAD) systems, especially those based on deep learning (DL), have shown promise in enhancing detection and classification of CRC by automatically learning features from medical images. However, publicly available colonoscopy datasets are scarce and often of limited quality, which hampers model performance. The purpose of this work is to investigate optimized deep learning approaches for colon cancer analysis, focusing on selecting effective optimization algorithms for CNN-based classification over multiple CRC datasets, with the broader aim of aiding early prediction to support timely treatment.

The related work surveys the evolution from classical machine learning with handcrafted features (e.g., SVM, k-NN, Random Forest) to deep learning methods that learn directly from raw images for CRC detection, classification, and polyp detection/localization. Multiple prior studies leveraged CNN architectures and transfer learning (e.g., VGG, ResNet, Inception, DenseNet) across imaging modalities including histopathology, endoscopy, colonoscopy, CT, MRI, PET, and OCT. Reported successes include high accuracies in classifying tumor vs. normal/benign tissues, anchor-free real-time polyp detectors, temporal modeling in colonoscopy videos, and segmentation of glandular structures. Reviews highlight DL’s strong performance across medical image analysis tasks and the ongoing need for robust datasets. While many works are CNN-based, few explicitly compare optimizers, and dataset limitations (especially colonoscopy) remain a bottleneck.

The study evaluates a CNN-based pipeline for colon cancer image classification while systematically comparing state-of-the-art optimizers: SGD, Adadelta, RMSprop, Adam, Adamax, and Nadam. CNN architecture: images are preprocessed (resizing, normalization, augmentation) and resized to 224×224×3. Standard CNN components are used: convolutional layers with learned filters for feature extraction, pooling layers for dimensionality reduction, dropout to mitigate overfitting, and fully connected layers with activation functions (e.g., ReLU, Softmax) for classification. Optimizers: mathematical formulations and update rules for SGD, Adadelta, RMSprop, Adam, Adamax, and Nadam are described and applied to train CNNs to assess their impact on performance. Datasets and splits: four datasets were used with an 80/10/10 split for train/validation/test: (DS_1) Colon tissue images (LC25000) with 500 images (250 benign, 250 adenocarcinoma), 768×768; (DS_2) Colonoscopy frames extracted from videos totaling 15 serrated adenomas, 21 hyperplastic lesions, and 40 adenomas, 768×576; (DS_3) Warwick-QU histology dataset with 165 images (74 benign, 91 malignant), 520×775; (DS_4) CRC-VAL-HE-7K histology with nine classes (ADI, BACK, DEB, LYM, MUC, MUS, NORM, STR, TUM), 224×224, from 50 patients. Image augmentation: to address limited data and improve generalization, extensive augmentation was applied. Parameters included rotation up to 30°, width/height shift 0.2, shear 0.2, zoom 0.2, horizontal/vertical flips, fill_mode='nearest'. DS_1 augmented in 10 ways to 5000 images; DS_3 in 16 ways to 2527 images; DS_2 and DS_4 also augmented (DS_4 had fewer augmentations due to larger size). Training setup: Implemented in Python using Keras with TensorFlow backend, trained on Google Colab GPU over 100 epochs. Evaluation metrics: precision, recall, F1-score, and accuracy were computed, including per-dataset results reported for each optimizer. DS_4 includes nine-class classification reflecting diverse tissue types relevant to CRC.

- Overall: CNN with Adam achieved the best average performance across four datasets, with an average accuracy of approximately 82%. - DS_1 (Colon; binary benign vs. adenocarcinoma): Top accuracies were Adadelta 0.96 and Adam 0.95. RMSprop 0.76; Nadam 0.79; SGD 0.67; Adamax 0.68. Precision/Recall/F1 for top methods: Adadelta (P=0.97, R=0.94, F1=0.96), Adam (P=0.91, R=0.95, F1=0.95). - DS_2 (Colonoscopy): Nadam performed best with Accuracy 0.79 (P=0.76, R=0.79, F1=0.76). Other methods achieved ≤0.69 accuracy. - DS_3 (Warwick-QU histology): Best accuracies were Adam and SGD at 0.76. Adadelta underperformed (Acc=0.59). Highest precision was Adamax (0.77), highest recall was SGD (0.78), highest F1 was Adam (0.72), indicating metric-specific strengths. - DS_4 (CRC-VAL-HE-7K, 9-class): Adam was best (Acc=0.90; P=0.89, R=0.87, F1=0.87). Adamax and Nadam also strong (Acc=0.87 each). SGD: Acc=0.73. - Learning curves: Adam, Adamax, Nadam typically showed smoother convergence and lower validation loss; some optimizers (e.g., RMSprop on DS_1, Adadelta on DS_4) showed instability or poorer convergence. - Data augmentation: Excessive augmentation, particularly in DS_3, may have introduced less informative samples, contributing to lower performance compared to datasets with more original informative images.

The study’s goal was to assess the potential of deep learning for colon cancer classification and to identify best practices around optimizer choice. Results show that while no single optimizer dominates every scenario, Adam consistently delivers strong or best performance across diverse datasets and tasks, achieving the top average accuracy and the highest performance on the largest, multiclass histology dataset (DS_4). On DS_2 (colonoscopy), Nadam was superior, highlighting that dataset characteristics and problem complexity can influence the optimal optimizer choice. The Warwick-QU results (DS_3) illustrate trade-offs across metrics and the risk that heavy augmentation of limited, heterogeneous data can degrade generalization. Learning curve analyses support these conclusions, suggesting more stable convergence for Adam-family optimizers in many settings. Collectively, these findings support the research hypothesis that optimized DL pipelines, particularly with carefully chosen optimizers and augmentation strategies, can enhance CRC image classification performance and provide a foundation for CAD tools to aid clinicians in early detection and screening.

The paper presents an empirical comparison of CNN optimizers for colon cancer image classification across four datasets, demonstrating that Adam generally provides the best or near-best performance, with an average accuracy of about 82% and top accuracy of 0.90 on the multiclass CRC-VAL-HE-7K dataset. For the Colon dataset (DS_1), both Adadelta (0.96) and Adam (0.95) performed strongly; Nadam was best on colonoscopy (DS_2). The work establishes performance benchmarks across datasets and underscores the impact of dataset quality and augmentation on outcomes. The authors emphasize that the classification metrics are not a direct clinical diagnosis; rather, the system can support screening and triage. Future work includes collecting and releasing more colonoscopy datasets, conducting external validation with oncologists/pathologists, refining augmentation strategies, and exploring improved architectures and multi-stage segmentation-classification pipelines for enhanced clinical utility.

- Dataset limitations: scarcity and uneven quality of publicly available colonoscopy datasets; small sample sizes in DS_2 and DS_3; potential class imbalance in DS_4 classes. - Augmentation effects: heavy augmentation (e.g., DS_3) may introduce less informative samples, hurting generalization and inflating training size without proportional performance gains. - Generalizability: results are based on specific datasets and may not directly transfer to other clinical settings without external validation. - Clinical applicability: evaluation uses image-level classification metrics; no prospective clinical validation or patient-level outcomes; models are not ready for diagnostic use without further validation. - Methodological scope: primary focus on optimizer comparison within a CNN framework; limited exploration of architecture search, segmentation pipelines, or domain adaptation that could further improve performance.