Captioning MRI Tumor Otak 3D Melalui Kata Kunci Spasial Dan Morfologis Berbasis Segmentasi Untuk Mengurangi Ketidakakuratan Anatomi

Hassan, Daffa Zimraan (2026) Captioning MRI Tumor Otak 3D Melalui Kata Kunci Spasial Dan Morfologis Berbasis Segmentasi Untuk Mengurangi Ketidakakuratan Anatomi. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221223-undergraduete_thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Magnetic Resonance Imaging (MRI) merupakan modalitas utama dalam diagnosis tumor otak. Namun, interpretasi manual yang menyita waktu rentan terhadap kelelahan visual dan variabilitas antar-pengamat. Medical Image Captioning (MIC) hadir untuk mengotomatisasi pelaporan medis ini. Akan tetapi, model MIC standar berarsitektur encoder-decoder sering mengalami halusinasi klinis, di mana teks yang dihasilkan lancar secara linguistik namun mengandung informasi faktual yang salah, khususnya terkait lateralisasi dan ukuran lesi. Untuk mengatasi permasalahan tersebut, penelitian ini mengusulkan kerangka kerja MIC dua tahap
(Generate-and-Refine) yang mengintegrasikan atribut klinis dari hasil segmentasi 3D sebagai batasan semantik tegas (hard semantic constraints). Model ini dievaluasi menggunakan dataset BraTS-GLI dan RadGenome-Brain MRI. Hasil eksperimen menunjukkan bahwa integrasi batasan segmentasi secara efektif mampu memitigasi halusinasi spasial. Akurasi penentuan lateralisasi tumor meningkat secara signifikan dari 29,59% (baseline) menjadi 76,92%. Selain itu, tahap penyempurnaan teks berhasil meningkatkan skor BLEU-4 sebesar 33,87% (dari 0,1051 menjadi 0,1407) serta nilai ekuivalensi semantik rata-rata F1 BERTScore dari 0,2943 menjadi 0,3517. Melalui pendekatan ini, sistem mampu menghasilkan laporan medis yang tidak hanya fasih secara linguistik, tetapi juga berlandaskan pada bukti visual yang akurat secara klinis sebagai purwarupa sistem bantuan diagnostik.
======================================================================================================================================
Magnetic Resonance Imaging (MRI) is the primary modality for brain tumor diagnosis. However, manual interpretation is time-consuming and prone to visual fatigue and inter
observer variability. Medical Image Captioning (MIC) offers a solution to automate medical reporting. Nevertheless, standard encoder-decoder MIC models frequently suffer from clinical hallucinations, generating linguistically fluent text with incorrect factual information, particularly regarding tumor laterality and lesion size. To address this issue, this study proposes a two-stage Generate-and-Refine MIC framework that integrates clinical attributes derived
from 3D segmentation as hard semantic constraints. The proposed model was evaluated using the BraTS-GLI and RadGenome-Brain MRI datasets. Experimental results demonstrate that integrating segmentation constraints effectively mitigate spatial hallucinations. Tumor laterality accuracy improved significantly from 29,59% (baseline) to 76,92%. Furthermore, the refinement stage improved the BLEU-4 score by 33,87% (from 0,1051 to 0,1407) and increased the mean semantic equivalence BERTScore F1 from 0,2943 to 0,3517. Through this approach, the system generates medical reports that are both linguistically fluent and clinically accurate, serving as a reliable diagnostic assistance prototype.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Image Captioning, Mitigasi Halusinasi Klinis, Model Bahasa-Visi Medis, Perbaikan Terpandu Segmentasi, Segmentasi Tumor Otak. Brain Tumor Segmentation, Clinical Hallucination Mitigation, Image Captioning, Medical Vision-Language, Segmentation-Guided Refinement.
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > T Technology (General) > T57.5 Data Processing T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Daffa Zimraan Hassan
Date Deposited:	24 Jul 2026 08:38
Last Modified:	24 Jul 2026 08:38
URI:	http://repository.its.ac.id/id/eprint/137335

Actions (login required)

View Item