Source-linked AI summary
GasHisSDB: A New Gastric Histopathology Image Dataset for Computer Aided Diagnosis of Gastric Cancer
Weiming Hu, Chen Li, Xiaoyan Li, Md Mamunur Rahaman, Jiquan Ma, Yong Zhang, Haoyuan Chen, Wanli Liu, Changhao Sun, Yudong Yao, Hongzan Sun, Marcin Grzegorzek
TL;DR
Manual gastric histopathology review is subjective and demanding, and evaluating classifiers requires sub-size image data. GasHisSDB provides three image-size sub-databases for classifier evaluation; deep learning outperformed classical methods in the reported comparisons, with VGG16 reaching 96.47% accuracy on Sub-database B.
Problem
Manual review is subjective and demanding, and sub-size image databases are needed to test image-classification methods.
Source: §1.1 · pp. 1–2Method
GasHisSDB provides three sub-databases of 160×160, 120×120, and 80×80 pixel images for evaluating classical and deep learning classifiers.
Source: §5 · p. 8Results
Deep learning models surpassed classical methods; VGG16 reached 96.47% accuracy on Sub-database B, versus 86.08% for classical RF using color histograms.
Source: §3.1 · p. 6, §4.2 · p. 8Takeaways & Limitations
GasHisSDB supports comparison of classifier discrimination performance on gastric histopathology sub-size images.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Background and Objective: Gastric cancer has turned out to be the fifth most common cancer globally, and early detection of gastric cancer is essential to save lives. Histopathological examination of gastric cancer is the gold standard for the diagnosis of gastric cancer. However, computer-aided diagnostic techniques are challenging to evaluate due to the scarcity of publicly available gastric histopathology image datasets. Methods: In this paper, a noble publicly available Gastric Histopathology Sub-size Image Database (GasHisSDB) is published to identify classifiers' performance. Specifically, two types of data are included: normal and abnormal, with a total of 245,196 tissue case images. In order to prove that the methods of different periods in the field of image classification have discrepancies on GasHisSDB, we select a variety of classifiers for evaluation. Seven classical machine learning classifiers, three Convolutional Neural Network classifiers, and a novel transformer-based classifier are selected for testing on image classification tasks. Results: This study performed extensive experiments using traditional machine learning and deep learning methods to prove that the methods of different periods have discrepancies on GasHisSDB. Traditional machine learning achieved the best accuracy rate of 86.08% and a minimum of just 41.12%. The best accuracy of deep learning reached 96.47% and the lowest was 86.21%. Accuracy rates vary significantly across classifiers. Conclusions: To the best of our knowledge, it is the first publicly available gastric cancer histopathology dataset containing a large number of images for weakly supervised learning. We believe that GasHisSDB can attract researchers to explore new algorithms for the automated diagnosis of gastric cancer, which can help physicians and patients in the clinical setting.
1. Introduction
Gastric cancer diagnosis relies on histopathology, but manual review is subjective and demanding, while sub-size image databases are needed to evaluate classifiers. GasHisSDB addresses this need with a large public dataset and classifier evaluations.
1.1. Research Background and Motivation
- Manual gastric-cancer review is subjective and demanding, motivating efficient computer-aided methods and sub-size images for testing classifier performance.The paper notes that whole-slide images are often cropped because of computer-performance constraints.Source: §1.1 · pp. 1–2
- GasHisSDB provides 245,196 sub-size images, expert-supported database preparation, and evaluations of traditional and deep learning classifiers.The database is released as open source for non-commercial purposes.Source: §1.1 · p. 2
1.2. Related work
- Existing gastric histopathology datasets include large images in small quantities or smaller collections, and studies often augment or preprocess them.The authors position GasHisSDB as a way to use existing resources while complementing their functionality.Source: §1.2 · p. 2
2.1. Database preparation
GasHisSDB is a collection of gastric histopathology sub-size images prepared for weakly supervised learning, with expert calibration and defined image-processing rules.
2.1. Database preparation
- 245,196 images are divided into 97,076 abnormal and 148,120 normal cases, with images stored as PNG and the preparation approved by Northeastern University's Ethics Committee.The database has three sub-databases: 160×160, 120×120, and 80×80 pixels.Source: §2.1 · p. 3
- Four pathologists supplied 600 labeled 2048×2048 source images; biomedical researchers prepared the sub-size images, which two experienced pathologists calibrated.The source images were intended for strong supervision, while the 245,196 sub-size images were prepared for weak supervision.Source: §2.1 · p. 3
- Normal sections were cropped directly, while abnormal sections were cropped around ground-truth cancer regions and filtered by cancer-area coverage.Images were randomly rotated and shuffled to reduce correlation among sub-size images from the same original image.Source: §2.1 · p. 3
2.2. Database description
GasHisSDB distinguishes normal tissue from cancer-containing abnormal tissue and uses H&E-stained pathological images.
2.2.1. Normal image
- Normal images contain no cancerous region and are identified by cells with little or no atypia, minimal mitosis, and regular single-layer arrangement.Normal images are made by cropping the whole image directly.Source: §2.2.1 · p. 3
2.2.2. Abnormal image
- Abnormal images contain gastric cancer, with irregular glandular structures and uneven, irregularly arranged cancer cells; cancer regions are selected using ground truth.The text describes both clear tumor–stroma boundaries and infiltrative cases where the boundary is unclear.Source: §2.2.2 · p. 4
2.2.3. Staining method
- H&E staining colors nuclei purplish blue and cytoplasm and extracellular matrix pink, helping distinguish cellular components and tissue structure.Compared with normal images, abnormal images show more widespread, messy purplish-blue areas.Source: §2.2.3 · p. 4
2.3. Methods of feature extraction
The study extracts color and texture descriptors from GasHisSDB images for machine-learning classification, using color histograms, LBP, and GLCM features.
2.3. Methods of feature extraction
- Three visual-feature families are extracted for classification: color histograms, local binary patterns (LBP), and gray-level co-occurrence matrix (GLCM) features.The paper uses color histograms for color and LBP and GLCM for texture.Source: §2.3 · p. 5, §2.3.2 · p. 5
2.3.1. Color histogram
- Color histograms represent global color proportions and are robust to rotation, shifts, and image-scale normalization, but omit local color distribution and spatial locations.They suit images that are difficult to segment automatically or do not require object-location information.Source: §2.3.1 · p. 5
2.3.2. Texture features
- LBP compares each pixel with eight neighbors for local texture, while GLCM captures paired-pixel brightness and location patterns and yields contrast, correlation, energy, and homogeneity measures.LBP is described as invariant to gray level and rotation.Source: §2.3.2 · p. 5
2.4. Methods of classification
The section outlines classical and deep-learning classifiers used to assess GasHisSDB. It describes the classical models, VGG16’s convolution design, and selected experiment settings.
2.4. Methods of classification
- The classifiers include Random Forest (RF), linear Support Vector Machine (linear SVM), VGG16, ResNet50, and ViT.RF and linear SVM are classical methods; VGG16, ResNet50, and ViT are deep-learning methods.Source: §2.4 · pp. 5–6
2.4.1. Classical machine learning methods
- Linear SVM classifies examples by mapping them into a space and assigning categories according to which side of the maximum gap they fall on.RF is an ensemble of decision trees that adds random attribute selection during training.Source: §2.4.1 · p. 6
2.4.2. Deep learning methods
- Deep learning is described as combining low-level features into more abstract representations through multiple hidden layers.The section situates this approach in artificial neural networks and distributed feature representations.Source: §2.4.2 · p. 6
- VGG16 replaces larger convolution kernels with stacks of 3×3 kernels to deepen the network while maintaining the same perceptual field.The described replacements are three 3×3 kernels instead of 7×7, and two 3×3 kernels instead of 5×5.Source: §2.4.2 · p. 6
- The classical comparisons use the same parameters, with 10 trees for RF and a linear kernel for SVM.The passage also mentions image sizes of 120 pixels and 80×80 pixels, but its surrounding description is incomplete.Source: §2.4.2 · p. 6
3.2. Results of deep learning methods
The experiments compare deep-learning classifiers across three GasHisSDB sub-databases under a shared training setup.
3.2. Results of deep learning methods
- The comparison evaluates classical and novel deep-learning methods on Sub-databases A, B, and C.Source: §3.2 · p. 6
- Each sub-database is split into training, validation, and test sets at a 4:4:2 ratio.Source: §3.2 · p. 6
- Models use a learning rate of 0.00002, batch size 32, and 100 training epochs.The passage says the results are shown in Table 4 but supplies no result values.Source: §3.2 · p. 6
3.3. Additional experiment
Because ViT had not fully converged after 100 epochs, the authors ran additional experiments, including a longer Sub-database A training run.
3.3. Additional experiment
- ViT had not fully converged after 100 epochs, prompting additional experiments reported as the last item for each sub-database in Table 4.Source: §3.3 · p. 6
- For Sub-database A, ViT ran for 400 epochs under the same parameter conditions, with training time kept similar to ResNet’s 100-epoch run.Source: §3.3 · p. 6
4. Discussion
Classical-method results vary by feature and classifier: RF performs best on color histograms, while linear SVM performs better on GLCM than on LBP. The discussion also outlines ResNet50 and ViT design choices.
4.1. Classical machine learning results discussion
- 85.99% is the RF accuracy on color-histogram features in Sub-database A, while linear SVM on those features scores below 50%.Source: §4.1 · p. 6
- 70.27% is the highest RF accuracy on LBP in Sub-database A, and linear SVM on LBP scores below 50%.Source: §4.1 · p. 6
- 66.50% is the linear SVM accuracy on GLCM features in Sub-database A; RF results on GLCM are similar to its LBP results.GLCM has four feature items per image, the fewest among the described features.Source: §4.1 · p. 6
- ResNet50 uses shortcut connections and a down-sampling block, with a 1×1 convolution added to the main branch of the latter [49].Source: §4.1 · p. 6
- ViT converts image chunks into linear embeddings that a Transformer processes for supervised image classification [50].Source: §4.1 · p. 6
3. Result
Classifier performance differed across GasHisSDB, with deep learning methods far ahead of classical methods in accuracy and experimental workload. The results also show how performance varied across sub-databases and classifiers.
3.1. Results of classical machine learning methods
- 86.08% on Sub-database B and 83.27% on C were RF's best color-histogram accuracies; RF also exceeded 70% on all three features in B.Source: §3.1 · pp. 6–7, Table 3 · p. 7
- 66.66% on Sub-database B's GLCM was linear SVM's only such result; in C, its classification result reached 60.81% and it still underperformed on LBP.Source: Table 3 · p. 7, §3.1 · p. 7
4.2. Deep learning results discussion
- In Sub-database A, ResNet50 reached 96.09%, the highest accuracy, while ViT used 1/4 of ResNet's training time and 1/3 of its model size.ViT accuracy was at least 86.21%, and its accuracy curve was still rising while its loss had not fully converged.Source: Table 4 · p. 7, §4.2 · p. 8
- VGG16 led at 96.47% in Sub-database B and 96.12% in C; ViT reached 89.44% and 90.23%, with C training time at 13.26% of VGG16's.Source: Table 4 · p. 7, §4.2 · p. 8
4.3. Additional experimental results discussion
- At matched ResNet50 training time, ViT reached 94.59% accuracy in Sub-database B and 94.57% in C; another experiment achieved 92.23%.The paper describes these results as reaching the general level of medical image classification and ViT as excellent in model size.Source: §4.3 · p. 8
4.4. Overall performance evaluation
- Across GasHisSDB, deep learning methods were far ahead of classical methods in image-classification accuracy and experimental workload, while classifier performance differed.Source: §4.4 · p. 8
5. Conclusion and futures works
GasHisSDB comprises three gastric histopathology sub-databases, each organized into normal and abnormal image folders. The paper evaluates classifiers across these databases and identifies future comparisons of newer methods as a research direction.
Conclusion and futures works
- GasHisSDB contains 160 × 160, 120 × 120, and 80 × 80 pixel sub-databases, each with normal and abnormal image folders of cropped, renamed, and shuffled images.The database is intended to discriminate classifier performance.Source: §5 · p. 8
- The evaluation tests five extracted features with seven classical classifiers across all three sub-databases, and compares CNNs and ViT using accuracy, model size, and training time.Additional training-time experiments seek ViT's best classification performance on GasHisSDB.Source: §5 · p. 8
- The authors plan to test newer image-classification methods on GasHisSDB to compare approaches and seek better practical methods.They state that creating the dataset means more image-classification methods can be used with it.Source: §5 · p. 8