Dual-Stream Approach For Acne Severity Grading with Ordinal Regression Methods

Raisa Aquila Zahra Kholiq, Prof. Dr.-Ing. Mhd. Reza M. I. Pulungan, S.Si., M.Sc.

2026 | Skripsi | ILMU KOMPUTER

Sistem penilaian keparahan jerawat otomatis sering memperlakukan tingkat keparahan sebagai kategori nominal independen, mengabaikan sifat progresif penyakit yang terurut. Penelitian ini mengusulkan arsitektur fusi dua aliran berbasis attention gate yang menggabungkan hitungan lesi lokal (YOLOv8m) dan fitur penampilan wajah global (EfficientNet-B4). Embedding yang difusikan dievaluasi pada tiga kepala klasifikasi: softmax+cross-entropy, CORAL, dan CORN, di bawah kondisi terkontrol menggunakan metrik Quadratic Weighted Kappa (QWK) dan Mean Absolute Error (MAE). Eksperimen pada set data ACNE04 menunjukkan bahwa CORN menghasilkan QWK rata-rata tertinggi (0,8609 ± 0,0061) dan kesalahan jarak terendah (MAE 0,2291±0,0117). Meski demikian, keunggulan CORN terhadap baseline softmax tidak signifikan secara statistik (p = 0,4922), sehingga keuntungan praktis utamanya terletak pada jaminan struktural konsistensi peringkat probabilitas kumulatif. Sebaliknya, performa CORAL menurun drastis akibat keterbatasan shared-weight constraint. Hasil ini menunjukkan bahwa efektivitas regresi ordinal pada penilaian medis tidak terjadi secara otomatis, melainkan bergantung pada kesesuaian asumsi struktural metode terhadap distribusi batas kelas data.

Automated acne severity grading systems often treat severity levels as independent nominal categories, introducing a fundamental misalignment with the ordered nature of disease progression. This work proposes an attention-gated dual stream fusion architecture that integrates local lesion counts (YOLOv8m) and global facial appearance features (EfficientNet-B4). The fused embedding is evaluated across three classification heads, namely softmax with cross-entropy, CORAL, and CORN, under controlled conditions using Quadratic Weighted Kappa (QWK) and Mean Absolute Error (MAE) metrics. Experiments on the ACNE04 dataset demonstrate that CORN achieves the highest mean QWK (0.8609 ± 0.0061) and the lowest distance error (MAE 0.2291 ± 0.0117). However, its performance margin over the softmax baseline is not statistically significant (p = 0.4922), meaning its primary practical advantage lies in its structural guarantee of rank-consistent cumulative probabilities. Conversely, CORAL exhibits substantial performance degradation due to the limitations of its shared-weight constraint. These findings indicate that ordinal regression methods do not automatically outperform standard classification on medical grading tasks; rather, their effectiveness relies on how closely their structural assumptions match the underlying class boundaries of the dataset.

Kata Kunci : Automated acne grading, Conditional Ordinal Regression for Neural Networks (CORN), Mean Absolute Error(MAE), Consistent Rank Logits(CORAL), Ordinal regression, Intermediate feature fusion