Question
How do conventional machine-learning and deep-learning approaches differ for liver CT segmentation?
Published paper · Medical imaging
Application of Machine Learning and Deep Learning in Medical Image Segmentation—an Example of Liver CT Images
2025 5th Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS)
Research summary
Scope
The paper compares SVM- and U-Net-based approaches across overlap metrics, speed, complexity, and interpretability.
Contribution
This was my first end-to-end research project, covering literature review, preprocessing, evaluation, and academic writing.
Abstract
In the realm of modern medicine, the process of liver image segmentation holds immense significance in facilitating precise diagnoses of hepatic diseases. However, this procedure is currently confronted with numerous challenges, including the challenges of achieving optimal accuracy and reliability in the results obtained. This study focuses on the application of machine learning and deep learning techniques to computed tomography (CT) liver images. The study utilizes support vector machine and convolutional neural network architectural models to assess the performance of the algorithms from diverse perspectives through experimental evaluations. The results of this investigation show the efficacy of a deep learning model in large-scale image segmentation, exhibiting superior accuracy and processing speed. In contrast, the machine learning algorithm demonstrates superiority in single-image processing. The innovation of the machine learning algorithm is achieved by incorporating an attention module and implementing multi-scale feature fusion, thereby boosting the model's practical performance and providing more reliable support for medical diagnosis.
Study design
The paper uses the public LiTS-ISBI2017 dataset, resizes images to 256×256, applies normalization and data augmentation, and discusses a 7:2:1 train/validation/test split. Dice coefficient and Jaccard index are used as overlap metrics.
Reported comparison
The published paper reports the following comparison. These values are presented as paper-reported results and are not an independent benchmark reproduced for this website.
| Metric | Deep learning (U-Net) | Conventional ML (SVM) |
|---|---|---|
| Dice | Above 0.90 | 0.75–0.80 |
| Jaccard | 0.85 | 0.70 |
| Single-image speed | 2–3 s (3D U-Net) | 0.5 s |
Paper-reported result: IEEE ACCTCS 2025 paper.
What this project established
This project established a first complete research workflow: defining a problem, organizing related work, describing preprocessing and evaluation, comparing model families, and preparing a conference manuscript. It also motivated my later shift toward more rigorous experimental protocols and reproducibility.
Scope and limitations
- The work is a broad comparative study rather than a new clinical segmentation system.
- Reported values should not be interpreted as a head-to-head reimplementation under a single audited codebase.
- Clinical use would require external validation, stronger baselines, and detailed reproducibility artifacts.