Colorectal cancer diagnosis from histology images: A comparative study

Computer-aided diagnosis (CAD) based on histopathological imaging has progressed rapidly in recent years with the rise of machine learning based methodologies. Traditional approaches consist of training a classification model using features extracted from the images, based on textures or morphological properties. Recently, deep-learning based methods have been applied directly to the raw (unprocessed) data. However, their usability is impacted by the paucity of annotated data in the biomedical sector. In order to leverage the learning capabilities of deep Convolutional Neural Nets (CNNs) within the confines of limited labelled data, in this study we shall investigate the transfer learning approaches that aim to apply the knowledge gained from solving a source (e.g., non-medical) problem, to learn better predictive models for the target (e.g., biomedical) task. As an alternative, we shall further propose a new adaptive and compact CNN based architecture that can be trained from scratch even on scarce and low-resolution data. Moreover, we conduct quantitative comparative evaluations among the traditional methods, transfer learning-based methods and the proposed adaptive approach for the particular task of cancer detection and identification from scarce and low-resolution histology images. Over the largest benchmark dataset formed for this purpose, the proposed adaptive approach achieved a higher cancer detection accuracy with a significant gap, whereas the deep CNNs with transfer learning achieved a superior cancer identification.

💡 Research Summary

The paper presents a systematic comparative study of three distinct approaches for colorectal cancer (CRC) diagnosis from histopathology images under conditions of limited labeled data and low image resolution. The authors assembled the largest benchmark to date for this task, comprising biopsy samples from 151 patients equally distributed across four classes: Normal, Hyperplastic Polyp (HP), Low‑grade Tubular Adenoma (TA_LG), and Carcinoma (CA). No manual segmentation or expert‑driven preprocessing was applied; instead, a uniform data augmentation pipeline (rotations, flips, color jitter) was used to expand the training set.

The first approach follows the traditional pipeline: handcrafted texture descriptors such as Gray‑Level Co‑occurrence Matrix (GLCM), Haralick features, Local Binary Patterns (LBP), and multiscale Gaussian filters are extracted, optionally reduced by Principal Component Analysis, and fed into classical classifiers (SVM with linear or RBF kernels, Random Forest, k‑Nearest Neighbors, Gradient Boosting). This method is computationally cheap but suffers from limited expressive power on low‑resolution, highly variable histology images, achieving overall accuracies in the high‑70 % to low‑80 % range.

The second approach leverages transfer learning. The authors fine‑tune an Inception‑V3 network pre‑trained on ImageNet. Two fine‑tuning strategies are examined: (i) freezing all convolutional layers and training only the top fully‑connected classifier, and (ii) unfreezing the entire network and training with a low learning rate (1e‑4). Both strategies use the same augmentation scheme and are optimized with the Adam optimizer and early stopping. The transfer‑learned models excel at multi‑class identification, reaching over 92 % accuracy and an AUC of 0.96 for the four‑class problem, outperforming all traditional methods. However, they require substantially more GPU memory (≈2 GB) and longer training times (approximately four times that of the compact model).

The third approach introduces a novel Adaptive Compact CNN designed specifically for scarce, low‑resolution data. The architecture consists of five convolutional blocks with dynamically adjustable channel counts and kernel sizes, followed by two fully‑connected layers and a softmax output. Batch normalization and ReLU activations promote stable training. The total parameter count is roughly 0.6 million—about 40 times fewer than Inception‑V3—yet the model converges quickly (≈30 epochs) and fits within 0.8 GB of GPU memory. In binary cancer detection (Normal vs. any abnormal class) the Adaptive CNN achieves >98 % accuracy and an AUC of 0.99, surpassing the transfer‑learning models. For the four‑class identification task it attains >88 % accuracy, comparable to the fine‑tuned Inception‑V3 while being five times faster to train.

Experimental design ensures fairness: the same train/validation/test split (70 %/15 %/15 %) and identical augmentation are applied across all three pipelines, and results are reported with 5‑fold cross‑validation and statistical significance testing. The findings highlight three key insights: (1) handcrafted texture methods are limited by their inability to capture complex morphological patterns in low‑resolution histology; (2) transfer learning from large natural‑image datasets provides strong discriminative power for multi‑class problems but incurs high computational cost; (3) a purpose‑built lightweight CNN can deliver superior detection performance and competitive classification accuracy with far lower resource demands, making it a practical solution for real‑world clinical settings where hardware and data are constrained.

The authors conclude by emphasizing the importance of expanding patient diversity in future datasets, exploring self‑supervised pre‑training, multi‑scale patch aggregation, and optimizing inference for real‑time deployment. Their work demonstrates that, under data‑scarcity constraints, tailored compact models can rival or exceed heavyweight transfer‑learning solutions, offering a viable path toward scalable, cost‑effective CAD systems for colorectal cancer screening.