Evaluasi Performa dan Stabilitas Pembelajaran Arsitektur ConvNeXt untuk Klasifikasi Perdarahan Otak pada Dataset CT Scan Terbatas

Tengku Syaid Farhan, Afiahayati, S.Kom., M.Cs., Ph.D.

2026 | Skripsi | ILMU KOMPUTER

Perdarahan intrakranial (Intracranial Hemorrhage, ICH) menyumbang sekitar 29% kejadian stroke dunia dan memerlukan diagnosis dalam golden hour, namun ketersediaan radiolog yang terbatas menjadi hambatan. Sistem kecerdasan buatan berpotensi menjadi alat bantu triase untuk mempercepat identifikasi kasus prioritas, bukan menggantikan peran radiolog. Namun, 89% penelitian klasifikasi ICH menggunakan dataset >5.000 sampel yang tidak sebanding dengan data yang dimiliki sebagian besar rumah sakit, dan tidak satupun melaporkan analisis stabilitas pembelajaran. Gap ini mendorong pertanyaan apakah arsitektur modern yang mengadopsi prinsip desain Vision Transformer, dengan komponen yang secara teoritis relevan untuk kondisi data terbatas, mampu memberikan performa yang lebih stabil dibanding arsitektur konvensional pada skala data realistis tingkat rumah sakit. Penelitian ini mengevaluasi ConvNeXt-Tiny, yang meskipun telah menunjukkan keberhasilan pada domain medis terbatas lain seperti deteksi leukoencephalopathy (Cernekova et al., 2024) dan klasifikasi Alzheimer berbasis MRI (Xia et al., 2025), belum pernah diuji pada klasifikasi biner ICH dengan dataset <500 xss=removed>

Intracranial hemorrhage (ICH) accounts for approximately 29% of global stroke cases and requires diagnosis within the golden hour, yet the limited availability of radiologists remains a barrier. Artificial intelligence systems have potential as triage aids to accelerate the identification of priority cases, not to replace radiologists. However, 89% of ICH classification studies use datasets of more than 5,000 samples, which is not comparable to the data available at most hospitals, and none report learning stability analysis. This gap raises the question of whether a modern architecture adopting Vision Transformer design principles, with components theoretically suited for limited-data conditions, can achieve more stable performance than conventional architectures at realistic hospital-scale data volumes. This study evaluates ConvNeXt-Tiny, which has demonstrated success on other limited medical imaging domains such as leukoencephalopathy detection (Cernekova et al., 2024) and Alzheimer’s classification from MRI (Xia et al., 2025), yet has never been tested on binary ICH classification with fewer than 500 samples. It is compared against ResNet-50, EfficientNet-B0, and LightGBM with radiomics features on the CQ500 dataset, with external validation on the RSNA dataset to measure domain shift. Evaluation covers four binary classification scenarios (Any-ICH, IPH, SDH, SAH) using AUC-ROC as the primary metric and learning stability analysis through perepoch AUC variance. ConvNeXt-Tiny outperforms LightGBM by 45.88% relatively, surpasses VGG16 (Castro et al., 2019) by 12.6?spite using a smaller training split, and maintains the smallest domain gap (?AUC = ?0.1865) alongside the highest external AUC 0.6949 on RSNA validation. ConvNeXt design components contribute to learning stability, with a peak AUC of 0.9141 on the SAH subtype. This study confirms that stability and generalization evaluation are clinical readiness indicators that cannot be overlooked.

Kata Kunci : Perdarahan intrakranial, ConvNeXt, klasifikasi CT, dataset terbatas, stabilitas pembelajaran, generalisasi lintas domain