Source-linked AI summary
AppendiGrade: An XAI-Enhanced Deep Learning Framework for Grading Appendicitis in Ultrasound with Gaussian Blur and Grad-CAM
Fahad Ahammed, Omar Faruq Shikdar, Navid Zaman, Md Tahsin, Md. Nawab Yousuf Ali, Golam Sorwar
TL;DR
Distinguishing complicated appendicitis from uncomplicated disease using ultrasound remains difficult, especially across major complication types. This paper evaluates optimized deep learning models with Grad-CAM explanations and reports 95.58% test accuracy for InceptionV3.
Problem
Ultrasound appendicitis studies remain limited by small datasets, minimal explainable AI, and binary or three-class classification, leaving major complicated forms insufficiently covered.
Method
The study trains four pretrained models on five-class ultrasound data, applies preprocessing and tuning, and uses Grad-CAM to highlight affected regions.
Results
95.58% test accuracy was achieved by InceptionV3 after optimization, compared with 69.21% initially.
Takeaways & Limitations
Grad-CAM provides visual details of model decisions, potentially helping physicians crosscheck highlighted areas of concern in ultrasound images.
Abstract
from arXiv · showhide
Appendicitis is one of the most common abdominal emergencies worldwide and requires prompt diagnosis and treatment to prevent life-threatening conditions. However, accurately differentiating complicated cases, such as perforation or abscess formation, from uncomplicated appendicitis remains a significant clinical challenge. Among other methods, ultrasound is a safer and more cost-efficient diagnostic technique because of the lack of radiation exposure. In this research, an advanced system capable of automatically detecting complicated appendicitis from ultrasound images was developed. A dataset consisting of 4679 ultrasound images with 5 classes, namely perforated, abscess, acute, appendicolith, and normal, was used for the proposed model training and testing. Four pretrained deep learning models, DenseNet201, InceptionV3, ConvNextTiny, and VGG19, have been employed for detecting and classifying complicated appendicitis. In the initial configuration, InceptionV3 achieved the second highest accuracy, with a value of 69.21%. Owing to suboptimal performance with raw images, further optimization techniques, including image preprocessing, hyperparameter tuning, model fine-tuning, and image sharpening, were applied. These enhancements significantly improved the model's performance, with an accuracy of 95.58% for InceptionV3. The model performance is then explained with gradient-weighted class activation mapping (Grad-CAM), which creates a heatmap of the regions responsible for the model's prediction of the infected areas. This could make crosschecking with experts much easier.
I. INTRODUCTION
Appendicitis is a prevalent abdominal emergency whose increasing incidence makes early, accurate diagnosis important, while conventional imaging can be time-consuming, expertise-dependent, and costly. This study develops a deep learning approach using ultrasound images to classify five appendicitis-related classes with pretrained models and image preprocessing.
- Clinical motivation: 63.55% was the worldwide increase in appendicitis incidence from 1990–2019, underscoring the importance of timely and accurate diagnosis.Traditional imaging methods are time-consuming, require expert radiologists, and involve significant training and cost.
- Study objective and data: Five classes—acute, perforated, abscess, appendicolith, and normal—were represented in a dataset of 4,679 ultrasound images.The study targets different types of complicated appendicitis using ultrasound data.
- Modeling approach: Four pretrained models—DenseNet201, InceptionV3, ConvNextTiny, and VGG19—were employed for appendicitis classification.Models were initially trained on raw images before further processing and optimization steps.
- Optimization approach: Gaussian blur, unsharp masking, and layer freezing were included among the study’s image preprocessing and model-development procedures.These procedures followed the initial training of the models on raw images.
II. LITERATURE REVIEW
Existing AI-based appendicitis diagnosis studies mainly use CT or MRI and report promising performance, but face dataset, interpretability, and retrospective-data limitations. Ultrasound research remains limited in scope, leaving inadequate coverage of complicated appendicitis forms.
- Existing AI-based diagnosis: AI-based automated diagnostic systems aim to reduce reliance on radiologists while providing instant results, but most current research focuses on CT and MRI.Byun achieved 85.5% accuracy but was limited to confirmed cases and CT imaging; Walid achieved an AUC of 0.868.
- Existing AI-based diagnosis: 72.5% accuracy was reported for AppendiXNet, while other studies reached 83.75% accuracy, 92.15% accuracy, and 88.2% accuracy under stated dataset or interpretability constraints.AppendiXNet used a dataset that was too small; Mijwil faced interpretability challenges; Pati lacked model interpretability; and Akbulut used retrospective data.
- Research gap: Ultrasound appendicitis studies are limited by small datasets, insufficient explainable AI, and binary or at most three-class classification, leaving major complicated forms undetected.The identified gap concerns detecting all major forms of complicated appendicitis.
III. RESEARCH METHODOLOGY
The methodology used a dataset covering four types of complicated appendicitis and progressively improved model training beyond raw images through preprocessing, layer freezing, and hyperparameter tuning. Grad-CAM was then applied to explain the model’s decisions.
- The dataset comprised four types of complicated appendicitis.
- Raw-image training produced insufficient results, motivating further optimization.
- Image preprocessing, layer freezing, and hyperparameter tuning were applied to improve performance.
- Grad-CAM was used to explain the model’s decisions.
A. Dataset Description
The study uses 4,679 prelabelled ultrasound images organized into five appendicitis-related classes. A certified physician reviewed and cross-checked the annotations for accuracy.
- Dataset composition: 4,679 prelabelled ultrasound images were obtained from Roboflow and categorized into five classes.The classes were abscess, appendicolith, perforated, acute, and normal.
- Dataset composition: The dataset was reported to have a balanced distribution across all five classes.Its size and diversity were also described as distinguishing features.
- Annotation verification: A certified physician reviewed and cross-checked the annotations to ensure label accuracy.
B. Dataset Pre-processing
The dataset was standardized, augmented, and enhanced with Gaussian blur followed by an unsharp mask to reduce noise and improve contrast. Stratified sampling divided the data into training, validation, and testing subsets while preserving class proportions.
- Image standardization and augmentation: Images were resized to 200 ∗200 pixels and normalized by scaling pixel values to facilitate faster convergence.Rotation and flipping were optionally applied to increase dataset diversity.
- Image standardization and augmentation: Raw-image training performed poorly because of speckle noise and poor tissue contrast.Gaussian blur was applied first to reduce noise, followed by an unsharp mask to sharpen images.
- Gaussian blur and image sharpening: The preprocessing combination improved edge detection by reducing noise and enhancing contrast.Gaussian blur produced a smoothed image, which was used with the original image to generate an edge-highlighting mask before sharpening.
- Dataset splitting: 70% of the dataset was allocated for training, 15% for validation, and 15% for testing through stratified sampling.This strategy preserved the class distribution across all newly created subsets.
C. Pretrained models
The study employed four pretrained models—DenseNet201, InceptionV3, ConvNextTiny, and VGG19—to evaluate different feature-extraction, efficiency, and architectural characteristics for ultrasound classification.
- Model selection: Four pretrained models were employed: DenseNet201, InceptionV3, ConvNextTiny, and VGG19.
- Model selection: DenseNet201 was selected for hierarchical feature extraction, InceptionV3 for multiscale features, ConvNextTiny for resource-efficient performance, and VGG19 as a simple-architecture benchmark.
D. Model fine tuning and hyperparameter optimization
The study fine-tuned the deep learning models by freezing selected layers and searching over layer-freezing combinations and training hyperparameters. For InceptionV3, tuning varied learning rate and L2 regularization to minimize validation loss.
- Model fine tuning: Fine-tuning froze 469 of 706 layers in DenseNet201 and 139 of 310 layers in InceptionV3.Layer-freezing combinations were tested through grid search.
- Hyperparameter optimization: Training hyperparameters were optimized through random search, using Adam as the optimizer.The searched hyperparameters were learning rate, patience, epochs, batch size, and optimizer.
- Hyperparameter optimization: InceptionV3 tuning evaluated learning rate and L2 regularization alongside validation loss to identify combinations that minimize validation loss.Table 1 reports the results of the hyperparameter-tuning iterations for InceptionV3.
E. Gradient-weighted class activation mapping
Grad-CAM was used as an explainable AI technique to improve interpretation of appendicitis classifications by highlighting image regions and features influencing model predictions.
- Method: Grad-CAM highlights regions of interest in medical images to improve model interpretability.The method was presented as an explainable AI approach for medical-image analysis.
- Method: The approach identifies features influencing predictions for different types of appendicitis.It improves transparency and helps explain factors contributing to the model’s decisions.
- Method: For an input image I and class c, Grad-CAM computes the gradient of class score y_c with respect to feature maps A_k.The gradients are calculated for a specific layer k.
- Method: The gradients are globally averaged across spatial locations.This averaging step follows the gradient computation for the selected feature maps.
IV. EXPERIMENTAL RESULTS
Optimization substantially improved ultrasound classification, with InceptionV3 emerging as the best-performing and most stable model. Its predictions were further supported by strong test results and Grad-CAM visualizations emphasizing diagnostically relevant regions.
- Performance Before Optimization: 72.08% accuracy was achieved by DenseNet201 before optimization, followed by InceptionV3 at 69.21%, ConvNextTiny at 61.03%, and VGG19 at 35.86%.Optimization was subsequently applied only to DenseNet201 and InceptionV3 to reduce computational cost.
- Performance After Optimization: 96% accuracy was reached by InceptionV3 after optimization, while precision, recall, and F1 score each rose above 90% from approximately 69%.The optimization used image sharpening, fine-tuning, and hyperparameter tuning.
- Model Stability and Generalizability: InceptionV3 showed smoother convergence and less unpredictable validation behavior than DenseNet201, whose fluctuations suggested overfitting and instability.Its superior test accuracy was attributed to multiscale feature extraction, fewer parameters, and improved handling of sharpened, noise-reduced images.
- Class-wise Performance: 96%, 98%, and 97% were the optimized precision, recall, and F1 score for abscess, rising from precision 61%; acute scores increased to 94%, 92%, and 93%.The reported acute baseline scores were 71%, 59%, and 65% for precision, recall, and F1 score, respectively.
- Test-set Evaluation: 1,537 out of 1,608 ultrasound images were correctly classified by optimized InceptionV3, with 7 abscess, 27 acute, 9 appendicolith, 14 normal, and 14 perforated images misclassified.The ROC analysis reported an AUC of 1.0 for four classes and 0.99 for the remaining class.
- Explainability Analysis: Grad-CAM heatmaps indicated that visually distinguishing regions of each complicated appendicitis type primarily drove correct predictions.The heatmap used red for greatest influence, yellow for moderate influence, and green to blue for low to very low importance.
V. CONCLUSION
The study found that default deep learning models applied to raw ultrasound images produced suboptimal, overfit results. Optimization techniques substantially improved InceptionV3 accuracy, while Grad-CAM was incorporated to improve interpretability.
- Default model setups on raw images yielded suboptimal results, with overfitting and significant disparities between training and testing accuracy.
- 95.38% InceptionV3 accuracy was achieved after applying multiple optimization techniques, increasing from 69.21%.
- Grad-CAM was incorporated to improve the interpretability of appendicitis detection predictions.