Get started
Work

LungColonCancerPred-NN: A Neural Network Approach for Lung and Colon Histopathology Prediction

Year: 2026
Type: preprint
Abstract: Lung and colorectal cancers are the first and second causes of cancer death worldwide, and their diagnosis rests on the reading of hematoxylin and eosin (H\&E) stained tissue. The LC25000 dataset is the most used public benchmark for automated classification of lung and colon histopathology, but its $25\,000$ images were produced by augmenting only $1\,250$ source images, so a random split places near copies of the same image in the training and test sets. We present LungColonCancerPred-NN, a compact triple-branch network ($5.8$M parameters) for five-class classification (colon adenocarcinoma, benign colon, lung adenocarcinoma, benign lung, and lung squamous cell carcinoma), and we evaluate it under a protocol designed to keep near copies apart. We combined LC25000 with a second public collection of the same classes ($40\,000$ files), removed $2\,076$ exact duplicates by MD5 hashing, grouped the remaining $37\,924$ images into $7\,795$ augmentation families using an EfficientNet-B0 embedding made invariant to the eight rotations and reflections of the square, and split the data $70/15/15$ by family. An audit of nearest-neighbor similarities confirmed the separation: the median similarity of a test image to its closest training image was $0.961$, against $0.998$ within the training set, and no test image exceeded the grouping threshold of $0.98$. The model shares one EfficientNet-B0 feature map between a convolutional pooling branch, a graph branch that passes messages between similar feature positions, and a Transformer encoder branch, and a sigmoid gate weights the three embeddings before classification. On the held-out test set of $5\,497$ images, LungColonCancerPred-NN reached an accuracy of $99.25\%$ (95\% bootstrap CI $99.02$-$99.49$), a macro-$F_1$ of $99.38\%$ ($99.19$-$99.57$), Cohen's $\kappa$ of $0.9905$ and a macro-AUC of $0.9997$. No malignant image was classified as benign, only two images were assigned to the wrong organ, and $33$ of the $41$ errors, squamous cell carcinomas were predicted as adenocarcinomas. This accuracy is higher than those reported by four recent (2025-2026) CNN, Vision Transformer, and CNN-Transformer methods on LC25000 ($97.00$-$99.11\%$) for the three lung classes alone, while our model handles five classes under a split grouped by augmentation family.
DOI: 10.21203/rs.3.rs-11473288/v1
Language: English
Open Access status: green
OpenAlex ID: W7225757713
Cites:
Cited by:
Related to: