GRM-081 - Evaluation of hand-crafted features with mask images on PanNuke dataset

Aug 5, 2026

Upload by Skyler

This study evaluates interpretable nucleus classification for histopathology by using eight machine learning models on hand-crafted feature sets derived from mask images. Built on the PanNuke dataset, 46 engineered features extracted from H&E images support pixel-accurate tissue morphology understanding. Bayesian optimization selects informative features and a weighted ensemble improves multi-class performance across neoplastic, inflammatory, connective, and epithelial nuclei. The resulting lightweight pipeline avoids deep segmentation architectures while enabling accurate, transparent diagnostic support.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Abstract

Introduction

Dataset

Nucleus Classification

Preprocessing

Methodology Workflow

Feature Extraction

Feature Selection

Model Training

Prediction

Results

Model Evaluation Metrics

Comparative Performance

Feature Selection Analysis

Conclusion

Publication Achievement

Author and Advisor

Evaluation of hand-crafted features with mask images obtained from PanNuke dataset using Bayesian optimization and machine learning models Conclusion Abstract GRM-081 Results Semantic image segmentation enabled precise, pixel-level understanding of tissue morphology, which was vital for tasks like cancer detection. In this study, we evaluated eight machine learning models using 46 carefully engineered features extracted from 5 , 179 H&E-stained images in the PanNuke dataset. By applying Bayesian optimization for feature selection and ensemble learning for classification, we developed interpretable and computationally efficient diagnostic tools aimed at supporting early disease identification in low-resource clinical settings. Introduction Although deep learning has advanced semantic segmentation, its black-box nature limited transparency especially in histopathology, where interpretability is critical. To address this, we extracted hand-crafted features capturing structural and textural characteristics from nuclei across 19 tissue types. We integrated advanced selection techniques with a weighted ensemble model, resulting in a transparent and accurate pipeline that eliminated the need for deep segmentation architectures. Dataset: Utilized the PanNuke dataset containing 5,179 H&E-stained tissue images , annotated with 161,739 nuclei across 19 tissue types. Nucleus Classification: Focused on four major nucleus categories: •Neoplastic (cancerous) •Inflammatory (immune-related) •Connective (structural support) •Epithelial (surface-lining) Preprocessing: •Applied Reinhard color normalization to standardize image color tones. •Used mask-based nucleus extraction to isolate individual nuclei. Methodology Workflow 1.Feature Extraction → Derive 46 features from nuclei mask images. 2.Feature Selection → Apply advanced techniques to select optimal features. 3.Model Training → Evaluate individual and ensemble models using selected features. 4.Prediction → Combine model predictions with weighted ensemble learning. Fig 1:Overall process of the feature extraction, feature selection, and weighted ensemble learning using seven machine learning models on 5 , 179 H&E images containing 19 tissue types of five cell nuclei. SCAN HERE !! Explore the Complete Research The evaluation of our research was conducted using eight machine learning models as Logistic Regression, Decision Tree, Random Forest, XGBoost , LightGBM , Support Vector Machine, Gradient Boosting, and a Weighted Ensemble on the PanNuke dataset. These images included 161 , 739 segmented nuclei categorized into neoplastic, inflammatory, connective, and epithelial types across 19 tissue types. To assess the classification performance of each model, we used four key metrics: precision, recall, F1-score, and accuracy. Among all models, the Weighted Ensemble achieved the highest overall performance across multiple nuclei types and tissue categories. XGBoost and SVM also demonstrated strong individual performance, particularly in identifying neoplastic and inflammatory nuclei. Meanwhile, Random Forest excelled in classifying connective nuclei in adrenal and breast tissues and epithelial nuclei in colon tissue, indicating its suitability for certain structural features captured by hand-crafted attributes. Across most tissues, models trained with feature subsets selected via Bayesian Optimization consistently outperformed those using unfiltered features, showcasing the impact of optimized feature selection on classification quality. COMPARISON OF DIFFERENT MACHINE LEARNING MODELS CLASSIFICATION In addition, the feature selection plots using ANOVA, Mutual Information, RFE, LASSO, and Random Forest Importance illustrated how Bayesian optimization contributed to identifying the most informative and efficient subsets. This combination of interpretable features and model ensembles resulted in reliable classification performance, even in mul

FAQs

What dataset and input representations were used in this study?

The study used the PanNuke dataset with 5,179 H&E-stained tissue images. It relied on mask-based nucleus extraction to derive 46 hand-crafted features from nuclei mask images.

How did Bayesian optimization improve the classification pipeline?

Bayesian optimization was used for feature selection, producing subsets that consistently outperformed using unfiltered features. Feature selection plots (ANOVA, Mutual Information, RFE, LASSO, Random Forest Importance) illustrate how informative and efficient feature subsets were identified.

Which model achieved the best overall performance and why?

The Weighted Ensemble achieved the highest overall performance across multiple nucleus types and tissue categories. XGBoost and SVM also showed strong individual results, while Random Forest performed particularly well for connective and epithelial nuclei in specific tissues.

Aug 5, 2026

Upload by Skyler