💻 computer science

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

The paper introduces LungColonCancerPred-NN, a compact triple-branch neural network that achieves state-of-the-art 99.25% accuracy in five-class lung and colon histopathology classification by addressing data leakage in the LC25000 benchmark through a rigorous augmentation-family-based evaluation protocol.

Original authors: outhman abbassi, Zeyneb Belganche, Wajih Rhalem, Nassim Kharmoum

Published 2026-11-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: outhman abbassi, Zeyneb Belganche, Wajih Rhalem, Nassim Kharmoum

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every year, millions of people are diagnosed with cancer, and two of the most common and deadly forms affect the lungs and the colon. When a doctor suspects these diseases, they rely on a pathologist to examine tiny slices of tissue under a microscope. These slices are stained with special dyes that turn cell nuclei purple and other parts pink, creating a complex map of the body's microscopic architecture. The pathologist's job is to read this map, distinguishing between healthy tissue and cancer, and then identifying exactly what kind of cancer it is. This is a critical task because the treatment for one type of lung cancer is often different from another, and the difference between a benign growth and a malignant one determines whether a patient needs surgery or chemotherapy. However, this process is slow, difficult, and depends entirely on the availability of highly trained human experts.

For years, scientists have tried to teach computers to read these microscopic maps, hoping to create a tool that can assist doctors or even work independently. They use large collections of digital images to train these computer programs, known as neural networks, to recognize patterns. But there is a hidden trap in how these collections are often built. Many of the images used for training are not unique photographs of different patients. Instead, they are created by taking a single original image and digitally rotating it, flipping it, or slightly shifting it to create many new versions. If a computer program is tested on an image that is just a flipped version of one it has already seen during training, it might get the answer right not because it understands the disease, but simply because it recognizes the picture. This flaw, known as data leakage, has made it difficult to know if these computer programs are truly learning to diagnose cancer or just memorizing the test.

A team of researchers at Mohammed V University in Morocco has now addressed this problem directly. They built a new computer system called LungColonCancerPred-NN and tested it using a much stricter method than has been used before. Instead of randomly shuffling the images, they carefully grouped them by their original source. They started with two large public collections of tissue images, totaling 40,000 files. First, they removed thousands of exact duplicates where the same digital file appeared twice. Then, they looked for the "families" of images that came from the same original patient sample. Using a smart system that understands how images look the same even when rotated or flipped, they grouped all these related images together. They then made sure that no group of related images was split between the training set and the test set. This meant the computer had to learn to recognize the disease from scratch, without ever seeing a near-copy of the test images during its training.

The computer system they built is surprisingly compact, containing about 5.8 million adjustable settings, which is small enough to run on a standard graphics card. It works by looking at the tissue images through three different "lenses" simultaneously. One lens focuses on the general texture and color of the cells, much like a human eye scanning for overall patterns. A second lens looks for specific shapes and groups of cells that are similar to each other, even if they are far apart in the image, allowing it to connect scattered parts of a tumor. The third lens pays attention to the exact location of every cell, understanding how the tissue is organized in space. A special mechanism then weighs the information from all three lenses for each image, deciding which view is most important before making a final decision.

When the researchers tested this system on 5,497 images that it had never seen before, the results were remarkably accurate. The computer correctly identified the tissue type in 99.25% of the cases. More importantly, it never made the most dangerous mistake: it never labeled a cancerous tissue as healthy. It did make a few errors, but they were mostly in confusing two specific types of lung cancer with each other, a challenge that even human pathologists sometimes face and require additional tests to resolve. The system also correctly identified which organ the tissue came from in nearly every case. When compared to other recent computer programs tested on the same dataset, this new system performed better, even though it was solving a harder problem by distinguishing five different types of tissue instead of just three.

The success of this work lies not just in the high accuracy, but in the way the test was designed. By ensuring that the test images were truly new and not just copies of the training data, the researchers proved that their system actually learned to recognize the biological features of cancer. They found that the computer's confidence in its answers was well-calibrated, meaning it knew when it was unsure. This approach offers a clearer path forward for developing artificial intelligence in medicine, showing that with careful testing, these tools can move beyond simple pattern matching to become reliable assistants for doctors. The researchers plan to test their system on a completely different set of images from a different hospital to see if it can handle real-world variations, a crucial step before such technology could ever be used to help patients.

Technical Summary: LungColonCancerPred-NN

Problem Statement
Lung and colorectal cancers are the leading causes of cancer mortality worldwide, with diagnosis relying heavily on the manual reading of hematoxylin and eosin (H&E) stained tissue sections. While deep learning offers a path toward computer-aided classification, the standard benchmark for this task, the LC25000 dataset, suffers from significant data leakage. The dataset's 25,000 images are generated by augmenting only 1,250 source images via rotations and flips. Consequently, random train-test splits often place near-identical copies of the same source image in both sets, allowing models to achieve inflated accuracy by memorizing training data rather than learning generalizable features. Furthermore, existing literature often reports high accuracy on subsets of the data (e.g., only lung classes) without addressing these leakage issues or the challenge of distinguishing between five distinct histological classes (colon adenocarcinoma, benign colon, lung adenocarcinoma, benign lung, and lung squamous cell carcinoma) simultaneously.

Methodology
The authors propose LungColonCancerPred-NN, a compact neural network architecture and a rigorous evaluation protocol designed to eliminate data leakage.

Data Protocol and Leakage Control:
- Data Source: The study combines the LC25000 dataset with a second public collection ("Cancer classification"), resulting in 40,000 initial files.
- Deduplication: Exact duplicates are removed using MD5 hashing, eliminating 2,076 files.
- Augmentation Families: To handle rotated and flipped copies (which differ in file bytes but represent the same source), the authors construct "augmentation families." They use an EfficientNet-B0 embedding made invariant to the eight symmetries of the square (rotations and reflections). Images are linked if their cosine similarity exceeds a threshold ( $\tau$ ) and they share the same label. A threshold of $\tau = 0.98$ was selected to ensure no family exceeds 80 images, resulting in 7,795 distinct families.
- Grouped Split: The data is split 70/15/15 by family, not by individual image, ensuring that no source image's augmented copies appear in both training and test sets.
- Audit: A nearest-neighbor audit confirms the split's integrity. The median similarity between a test image and its closest training image is 0.961, significantly lower than the 0.998 similarity found within the training set. No test image exceeds the grouping threshold of 0.98.
Architecture (LungColonCancerPred-NN):
- Backbone: The model utilizes a single EfficientNet-B0 feature map (projected to 256 channels) as a shared input for three parallel branches.
- Convolutional Branch: Performs global average pooling to capture local texture features (nuclear size, chromatin, cytoplasm color).
- Graph Branch: Constructs a graph where nodes are feature positions (7x7 grid). Edges connect positions based on feature similarity (k-NN with $k=8$ ) rather than spatial proximity. This allows regions with similar histological structures (e.g., scattered tumor nests) to exchange information regardless of distance.
- Transformer Branch: Uses a learnable class token and positional embeddings with two pre-norm Transformer encoder layers to model global spatial relationships and the overall tissue layout.
- Gated Fusion: A sigmoid gate dynamically weights the embeddings from the three branches based on the input image before they are concatenated and passed to a classifier.
- Training Strategy: The model (5.8M parameters) is trained in two phases: first with the backbone frozen, then end-to-end. It employs label smoothing ( $\epsilon=0.08$ ), stochastic weight averaging (SWA), and three-view test-time augmentation (identity, horizontal flip, vertical flip).

Key Contributions

Leakage-Controlled Evaluation Protocol: The paper introduces a robust protocol for LC25000 and derived collections involving MD5 deduplication, D4-invariant family grouping, and a nearest-neighbor audit to verify that test images are distinct from training images.
Triple-Branch Architecture: A novel network design that fuses convolutional texture, graph-based feature similarity, and Transformer-based spatial attention on a single EfficientNet-B0 backbone, weighted by a learnable sigmoid gate.
Comprehensive Evaluation: A detailed assessment on a 5-class task (including both organs and malignancy types) using a strict grouped split, providing confidence intervals and error analysis that previous studies on LC25000 often lack.

Results
Evaluated on a held-out test set of 5,497 images (from 1,114 families):

Performance: The model achieved 99.25% accuracy (95% CI: 99.02–99.49), a macro-F1 of 99.38%, Cohen's $\kappa$ of 0.9905, and a macro-AUC of 0.9997.
Error Analysis:
- Malignancy Detection: 100% sensitivity; no malignant image was misclassified as benign.
- Organ Classification: 99.96% accuracy; only 2 images were assigned to the wrong organ.
- Specific Errors: Of the 41 total errors, 33 were lung squamous cell carcinomas misclassified as lung adenocarcinomas. This reflects the difficulty in distinguishing poorly differentiated tumors that share solid growth patterns, a challenge that often requires immunohistochemistry in clinical practice.
Comparison: The model outperformed four recent methods (2025–2026) that reported accuracies between 97.00% and 99.11% on the three lung classes alone. The authors note that those studies used random splits on fewer classes, whereas LungColonCancerPred-NN handles five classes under a leakage-controlled grouped split.

Significance and Claims
The paper claims that its primary significance lies in addressing the "inflated results" common in medical imaging benchmarks due to data leakage. By demonstrating that high accuracy (99.25%) can be achieved even when near-copies are strictly separated, the authors argue that their model's performance is robust and not an artifact of memorization.

The authors emphasize that the evaluation protocol (hashing, family grouping, and audit) is architecture-agnostic and should be adopted for any future work on LC25000. They modestly note that while their model achieves state-of-the-art results on this benchmark, the remaining errors (specifically between lung adenocarcinoma and squamous cell carcinoma) mirror the difficulties faced by human pathologists, suggesting that such cases may inherently require additional diagnostic tools beyond H&E staining. The paper concludes by stating the model is ready for testing on external datasets like LungHist700 to further validate generalizability.

Technical Summary: LungColonCancerPred-NN

More like this