Pengembangan Model untuk Prediksi Tags Abnormality dan Pembangkitan Laporan Medis Radiologi pada Dadat X-Ray Paru-Paru

Tsaniya, Hilya (2025) Pengembangan Model untuk Prediksi Tags Abnormality dan Pembangkitan Laporan Medis Radiologi pada Dadat X-Ray Paru-Paru. Doctoral thesis, Institute Teknologi Sepuluh Nopember.

Text 7025222004-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Pengembangan sistem medical image captioning bertujuan untuk membantu ahli radiologi dalam penulisan laporan medis yang terdiri dari label abnormalitas, diagnosis, dan observasi citra. Tantangan utama dalam bidang ini meliputi penurunan kualitas citra medis akibat proses akuisisi, ketimpangan distribusi label abnormal pada dataset publik, serta rendahnya koherensi dan kesesuaian semantik antara laporan prediksi dan laporan referensi. Penelitian ini mengusulkan pengembangan model pembelajaran yang mampu melakukan prediksi abnormalitas melalui pendekatan Multi-label Classification serta menghasilkan laporan medis terstruktur secara paralel menggunakan arsitektur encoder-decoder berbasis transformer. Proses dimulai dengan peningkatan kualitas citra menggunakan kombinasi Adaptive Gamma correction (AGC) dan Contrast Limited Adaptive Histogram Equalization (CLAHE), dilanjutkan dengan ekstraksi fitur citra X-ray dan teks medis. Komponen utama meliputi komparasi metode classifier untuk prediksi abnormality tags dengan penanganan data tidak seimbang menggunakan Remedial sampling, serta integrasi metode transformer untuk menyusun laporan medis berdasarkan hasil prediksi label dan observasi klinis. Selain evaluasi konvensional seperti BLEU, penelitian ini juga mengusulkan metode evaluasi berbasis entitas medis dengan memanfaatkan pemetaan ke ontologi standar (RadLex/UMLS) dan pengukuran kesamaan semantik menggunakan embedding BioBERT. Hasil menunjukkan bahwa kombinasi AGC dan CLAHE mampu meningkatkan kualitas citra secara kuantitatif dan kualitatif, sedangkan penggunaan Knowledge Distillation dalam proses ekstraksi fitur dan prediksi label abnormal menghasilkan performa lebih baik dibandingkan model pembanding. Evaluasi berbasis entitas juga memberikan penilaian yang lebih bermakna secara klinis dibandingkan metrik berbasis teks semata.
================================================================================================================================
The development of a medical image captioning system aims to assist radiologists in writing medical reports consisting of abnormality labels, diagnoses, and image observations. Key challenges in this area include the degradation of medical image quality due to the acquisition process, the unequal distribution of abnormality labels in public datasets, and the low coherence and semantic match between predicted reports and reference reports. This research proposes the development of a learning model capable of predicting abnormalities through a multi-label classification approach and generating structured medical reports in parallel using a transformer-based encoder-decoder architecture. The process begins with image quality improvement using a combination of Adaptive Gamma Correction (AGC) and Contrast Limited Adaptive Histogram Equalization (CLAHE), followed by feature extraction of X-ray images and medical text. Key components include a comparison of classifier methods for predicting abnormality tags with remedial sampling for handling imbalanced data, and an integration of transformer methods to compile medical reports based on predicted labels and clinical observations. In addition to conventional evaluation methods such as BLEU, this research also proposes a medical entity-based evaluation method utilizing mapping to standard ontologies (RadLex/UMLS) and semantic similarity measurement using BioBERT embedding. The results show that the combination of AGC and CLAHE can improve image quality quantitatively and qualitatively, while the use of Knowledge Distillation in the feature extraction and abnormal label prediction processes produces better performance than the comparison models. Entity-based evaluation also provides more clinically meaningful assessments than solely text-based metrics

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Medical image captioning, Image Enhancement, Multi-label classification, Attention, Transformer, Medical image captioning, Image Enhancement, Multi-label classification, Attention, Transformer
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Hilya Tsaniya Ismet
Date Deposited:	05 Aug 2025 03:12
Last Modified:	11 Aug 2025 02:07
URI:	http://repository.its.ac.id/id/eprint/126350

Actions (login required)

View Item