Comparative analysis of random forest and deep learning approaches for automated acute lymphoblastic leukemia detection using morphologicaland textural features
Acute Lymphoblastic Leukemia (ALL) is a type of blood cancer that requires early and accurate detection for effective treatment. Current diagnostic approaches face significant challenges including time-consuming manual examination, inter-observervariability, and difficulty in balancing sensitivity...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Informatics Department, Engineering Faculty
2025-07-01
|
| Series: | Jurnal Ilmiah Kursor: Menuju Solusi Teknologi Informasi |
| Subjects: | |
| Online Access: | https://kursorjournal.org/index.php/kursor/article/view/427 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Acute Lymphoblastic Leukemia (ALL) is a type of blood cancer that requires early and accurate detection for effective treatment. Current diagnostic approaches face significant challenges including time-consuming manual examination, inter-observervariability, and difficulty in balancing sensitivity with specificity. This study aims to develop and compare two automated ALL detection methodologies to overcome these limitations. We propose: (1) a Random Forest classifier using carefully engineered morphological and textural features, and (2) a Convolutional Neural Network (CNN)architecture for automated feature learning from microscopic blood cell images. Using 10,661 images from the ALL Challenge dataset, we evaluated both approaches on training (70%), validation (15%), and test (15%) sets. Feature importance analysis revealed cell area (10.71%), energy (10.67%), and skewness (10.50%) as the mostsignificant discriminative features. The Random Forest achieved 85% accuracy withnotable sensitivity for ALL detection (93%), while the deep learning approachdemonstrated superior performance with 87% accuracy and better false positive control(27.50% vs. 35.76%). Our comparative analysis shows that while both methodsdemonstrate clinical viability for automated ALL screening, the deep learning approachoffers advantages in reducing false positives while maintaining high detectionsensitivity. This research contributes to the advancement of computer-aideddiagnostic tools that can support pathologists in early ALL detection,potentially reducingdiagnostic time and improving consistency.
|
|---|---|
| ISSN: | 0216-0544 2301-6914 |