Colorectal cancer unmasked: A synergistic AI framework for Hyper-granular image dissection, precision segmentation, and automated diagnosis

Abstract

Colorectal cancer (CRC) is the second most common cause of cancer-related mortality worldwide, underscoring the necessity for computer-aided diagnosis (CADx) systems that are interpretable, accurate, and robust. This study presents a practical CADx system that combines Vision Transformers (ViTs) and DeepLabV3 + to accurately identify and segment colorectal lesions in colonoscopy images.The system addresses class balance and real-world complexity with PCA-based dimensionality reduction, data augmentation, and strategic preprocessing using recently curated CKHK-22 dataset comprising more than 14,000 annotated images of CVC-ClinicDB, Kvasir-2, and Hyper-Kvasir. ViT, ResNet-50, DenseNet-201, and VGG-16 were used to quantify classification performance. ViT achieved best-in-class accuracy (97%), F1-score (0.95), and AUC (92%) in test data. The DeepLabV3 + achieved segmentation state-of-the-art for tasks of localisation with 0.88 Dice Coefficient and 0.71 Intersection over Union (IoU), ensuring sharp delineation of areas that are malignant. The CADx system accommodates real-time inference and served through Google Cloud for information that accommodates scalable clinical implementation. The image-level segmentation effectiveness is evidenced by comparison with visual overlay and expert-manually deliminated masks, and its precision is illustrated by computation of precision, recall, F1-score, and AUC. The hybrid strategy not only outperforms traditional CNN strategies but also overcomes important clinical needs such as detection early, balance of highly disparate classes, and clear explanation. The proposed ViT-DeepLabV3 + system establishes a basis for advanced AI support to colorectal diagnosis by utilizing self-attention strategies and learning with different scales of context. The system offers a high-capacity, reproducible computerised colorectal cancer screening and monitoring solution and can be best deployed where resources are scarce, and it can be highly desirable for clinical deployment.

Keywords: Automatic segmentation; DeepLabV3+; Diagnosis of colorectal cancer; Hyper-granular image analysis; Vision transformers.

Fig. 1

The CADx architecture of CNNs and transformers with DeepLabV3 +

Fig. 2

The 10 classes CKHK-22 mixed dataset sample images

Fig. 3

CKHK-22 mixed dataset pre-processed sample images

Fig. 4

The train and test split of the CKHK-22 dataset

Fig. 5

PCA applied on CKHK-22 colonoscopy images

Fig. 6

Train vs test data for haemorrhoids and polyps classes

Fig. 7

The CNN architecture of the ResNet-50

Fig. 8

The map of features extracted by CNN ResNet-50

Fig. 9

The CNN architecture of the DenseNet-201

Fig. 10

A map of the features extracted by CNN DenseNet-201

Fig. 11

The CNN architecture of the VGG16

Fig. 12

A map of the features extracted by CNN VGG16

Fig. 13

The network architecture of the vision transformer

Fig. 14

The attention map extracted by the vision transformer network

Fig. 15

DeepLabV3 +’s illustrative architecture for colorectal cancer recognition

Fig. 16

The CADx flow diagram for the detection of colorectal cancer

Fig. 17

Hyperparameters for DeepLabV3 +, ViT, and CNN

Fig. 18

Comparison of the training and testing accuracy and AUC of CNNs and ViT

Fig. 19

The performance metrics of the models in CADx

Fig. 20

Accuracy of vision transformers and training loss

Fig. 21

The confusion matrix for the classification of two classes

Fig. 22

The Receiver Operating Characteristic (ROC) curve for classification

Fig. 23

The graphs depicting the accuracy and loss of the DeepLabV3 + model

Fig. 24

The detection of malignant Polyps with DeepLabV3 +

Fig. 25

The CADx system’s overarching architecture