AI Blends Deep and Handcrafted Features to Spot Lung and Colon Cancer with 99.4% Accuracy

Every year, more than two million people worldwide are diagnosed with lung cancer and nearly as many with colon cancer, according to World Health Organization figures cited by the researchers. For most of these patients, the decisive moment in their diagnosis comes not in a scanning suite but at a microscope, where a pathologist examines thin slices of stained tissue and searches for the subtle architectural fingerprints of malignancy. This manual process remains the gold standard, but it is slow, subjective, and heavily dependent on the experience of the observer. A new study published in Discover Artificial Intelligence proposes a way to lighten that burden: an artificial intelligence framework that fuses multiple kinds of image features and combines several classifiers into a single, highly accurate, and interpretable decision-making system.

The research team, led by Onkar Singh of Manipal University Jaipur together with colleagues from several Indian institutions, built their framework around a simple but powerful observation. Convolutional neural networks, the workhorses of modern medical image analysis, are superb at capturing global, high-level patterns in an image, but they can miss the fine-grained local texture cues that pathologists rely on when distinguishing, say, a benign gland from an early adenocarcinoma. Handcrafted descriptors such as Histogram of Oriented Gradients and Local Binary Patterns, by contrast, excel at encoding exactly those edges, gradients, and micro-textures, yet they are fragile when staining conditions or imaging equipment change. The researchers reasoned that the two approaches are complementary, and that a hybrid representation combining both should outperform either alone.

The technical pipeline works in three phases. First, each histopathology image is passed through a VGG16 network pretrained on ImageNet, with its convolutional layers frozen, and a 512-dimensional deep feature vector is extracted by global average pooling of the final convolutional layer. In parallel, the grayscale version of the image is processed to yield 6,084 HOG features describing edge and shape structure, plus a compact 10-dimensional LBP descriptor summarizing local texture. Concatenating these three streams produces a single 6,606-dimensional hybrid vector per image. Notably, the handcrafted features are computed on grayscale images to make them robust to stain variation, while the CNN branch retains color-related discriminative information.

A vector of that size is unwieldy. High-dimensional feature spaces demand heavy computation, invite overfitting, and often contain redundant or weakly informative components. To address this, the team applied Incremental Principal Component Analysis, a variant of PCA that processes data in mini-batches of 500 samples and is therefore scalable to large datasets. A scree plot of cumulative explained variance showed the curve flattening at around 100 components, so the researchers compressed the 6,606-dimensional vector down to just 100 principal components while preserving 95 percent of the variance. This single step made training faster and more stable, and, as the ablation experiments later revealed, actually improved accuracy rather than sacrificing it.

On the compressed features, four deliberately heterogeneous base classifiers were trained: a Random Forest of 100 decision trees, which handles non-linear boundaries and noisy features; a Support Vector Machine with a radial basis function kernel, which excels at margin-based separation; a Logistic Regression model, which is simple and well calibrated; and a compact multi-layer perceptron trained for 20 epochs with the Adam optimizer. Each classifier outputs class probabilities across the five histological categories of the benchmark dataset. These probability vectors are concatenated into a 20-dimensional meta-feature vector, which feeds a second-level logistic regression meta-learner. This stacking architecture, rooted in Wolpert’s classic work on stacked generalization, allows the final model to learn how best to weigh the strengths and weaknesses of each base classifier, reducing individual model bias.

Evaluated on the publicly available LC25000 dataset of 25,000 lung and colon histopathology images, the framework delivered striking results. The stacking ensemble achieved 99.38 percent accuracy, 99.40 percent precision, 99.35 percent recall, and an F1-score of 99.38 percent, surpassing every individual base model, of which the Random Forest was the strongest at 98.71 percent. The ensemble correctly classified 4,971 of 5,000 test images, with the colon categories essentially flawless and only minor residual confusion between lung adenocarcinoma and lung squamous cell carcinoma, the two visually similar subtypes. Five-fold cross-validation yielded a consistent 98.88 percent accuracy, and the receiver operating characteristic analysis produced an area under the curve of 1.00, indicating near-perfect class separation. Inference takes roughly 248 milliseconds per image on a workstation with an NVIDIA RTX 3090 GPU, a moderate cost the authors argue is justified by the accuracy gains.

What distinguishes this work from many high-performing black-box models is its commitment to explainability, a critical requirement for clinical adoption. The team applied SHAP, or SHapley Additive exPlanations, directly to the logistic regression meta-classifier, quantifying how much each base learner’s probability outputs contribute to the final decision. The analysis revealed that the Support Vector Machine’s predictions carried the largest overall influence across all five classes, with mean absolute SHAP contributions ranging from 1.685 to 2.415 depending on the class, while the Random Forest provided the next most important signal. Because each meta-feature corresponds to a concrete probability from a named classifier, the explanation has an explicit, human-readable meaning at the ensemble level.

To complement this decision-level explanation with spatial information, the researchers also performed occlusion sensitivity analysis, a model-agnostic technique in which local regions of an input image are systematically masked and the entire inference pipeline is recomputed. The drop in the predicted-class probability after masking indicates how important that region was to the decision. The resulting heatmaps showed that the ensemble focuses non-uniformly on tissue regions consistent with pathological features rather than on arbitrary artifacts, offering pathologists a visual map of where the model is looking. Ablation studies reinforced the design logic: CNN features alone reached 98.20 percent accuracy, HOG alone 97.65 percent, and LBP alone 95.10 percent, while the full fused representation compressed by IPCA delivered the best result of 99.38 percent, confirming that no single feature modality is sufficient on its own.

The authors are candid about the limitations. McNemar’s statistical tests confirmed significant improvements over the Random Forest, Logistic Regression, and the custom neural network, but the advantage over the Support Vector Machine alone was not statistically significant. When the framework was tested on the smaller and more challenging LungHist700 dataset, accuracy dropped to 74.10 percent, reflecting the limited sample size, the high visual similarity between adenocarcinoma and squamous cell carcinoma, and the substantial variability in staining, magnification, and tumor heterogeneity found in real-world material. The team also notes that LC25000, while balanced and curated, does not fully capture the messiness of clinical practice, and that the multi-model pipeline carries a higher computational cost than simpler single-model approaches.

Looking forward, the researchers outline several directions: validating the system on multi-center datasets with diverse staining protocols, developing lightweight fusion and ensemble methods to shorten inference time, incorporating domain-specific morphological cues such as nuclear atypia and gland shape directly into feature extraction, and exploring semi-supervised and active learning to reduce dependence on large labeled datasets. The long-term ambition is a practical decision-support tool that can sit beside pathologists in routine diagnostic workflows. For now, the study offers a compelling demonstration that the path to trustworthy medical AI may not run through ever-larger neural networks alone, but through the intelligent combination of old and new, with transparency built in from the start.

Subject of Research: Explainable ensemble learning for multi-modal feature fusion in lung and colon cancer histopathology image classification

Article Title: Multi-modal feature fusion and classification of histopathological images using ensemble learning

Article References: Singh, O., Pandey, C., Singh, K. K., Singh, V. P., & Soni, P. K. (2026). Multi-modal feature fusion and classification of histopathological images using ensemble learning. Discover Artificial Intelligence, 6(1), Article 1391. https://doi.org/10.1007/s44163-026-02277-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02277-x

Keywords: histopathology, lung cancer, colon cancer, ensemble learning, feature fusion, VGG16, HOG, LBP, incremental PCA, SHAP, explainable AI, LC25000