Rochmawati, Naim (2026) Pendekatan Deep Learning Multifaset untuk Deteksi Lesi Peritoneal Carcinomatosis pada Citra Laparoskopi. Doctoral thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 7025231012-Doctoral.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
7025231012-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy |
Abstract
Peritoneal carcinomatosis (PC) adalah kanker metastatik yang menyerang lapisan peritoneum di rongga perut dan berdampak besar terhadap prognosis pasien. Keberhasilan terapi seperti cytoreductive surgery (CRS) dan hyperthermic intraperitoneal chemotherapy (HIPEC) sangat bergantung pada akurasi deteksi lesi melalui citra laparoskopi. Namun, tantangan utama dalam deteksi berbasis video laparoskopi adalah keberadaan lesi kecil dengan kontras rendah, lesi besar yang menyatu dengan permukaan organ, serta citra yang mengalami gangguan visual akibat gerakan kamera dan faktor intraoperatif lainnya.
Penelitian ini mengadopsi pendekatan deep learning multifaset, yaitu strategi yang mencakup beberapa aspek sekaligus untuk mengatasi tantangan deteksi lesi PC, mulai dari evaluasi strategi pelatihan model, modifikasi arsitektur untuk meningkatkan kemampuan deteksi lesi kecil, hingga pengembangan pendekatan awal analisis statistik lesi pada video laparoskopi. Dataset yang digunakan merupakan citra laparoskopi nyata yang telah dianotasi oleh ahli dari Institut de Cancérologie de l'Ouest (ICO), Perancis. Tiga pendekatan dievaluasi dalam rangka meningkatkan performa deteksi menggunakan YOLO, yaitu: (1) perbandingan pelatihan dari awal (scratch) dengan transfer learning menggunakan pre-trained weights; (2) modifikasi arsitektur YOLOv8 melalui integrasi Convolutional Block Attention Module (CBAM) dan penambahan layer P2; serta (3) pengembangan sistem analisis video berbasis seleksi frame unik menggunakan deep feature extraction ResNet50 dan cosine similarity untuk menghasilkan statistik jumlah lesi terdeteksi dan estimasi luasan lesi berdasarkan hasil segmentasi.
Hasil eksperimen menunjukkan bahwa pendekatan transfer learning menggunakan pre-trained weights memberikan peningkatan performa deteksi yang signifikan dibandingkan pelatihan dari awal, dengan YOLOv8s berbobot pre-trained menghasilkan mAP50 sebesar 0,871 dan F1-Score sebesar 0,840, sedangkan pelatihan dari awal (scratch) hanya mencapai mAP50 sebesar 0,799 pada YOLOv8l. Sementara itu, modifikasi arsitektur YOLOv8 melalui integrasi CBAM dan penambahan layer P2 meningkatkan kemampuan model dalam mendeteksi lesi kecil, dengan presisi sebesar 0,887, recall sebesar 0,769, mAP50 sebesar 0,861, dan F1-Score sebesar 0,824. Studi ablasi menunjukkan bahwa setiap komponen, yaitu CBAM, layer P2, dan fungsi kerugian CIoU, memberikan kontribusi terhadap peningkatan performa model. Selain itu, pendekatan seleksi frame unik mampu mengurangi redundansi frame pada video laparoskopi sehingga proses analisis lesi dapat dilakukan secara lebih efisien. Sistem yang dikembangkan menghasilkan statistik jumlah lesi terdeteksi dan estimasi luasan lesi berdasarkan frame hasil seleksi sebagai langkah awal menuju pengembangan sistem analisis beban lesi berbasis video laparoskopi.
Penelitian ini memberikan kontribusi dalam pengembangan sistem deteksi dan analisis lesi otomatis berbasis visi komputer pada citra laparoskopi, serta membuka peluang untuk pengembangan sistem analisis lesi yang lebih komprehensif guna mendukung pengambilan keputusan klinis pada bedah onkologi di masa depan.
=============================================================================================================================================
Peritoneal carcinomatosis (PC) is a metastatic cancer that affects the peritoneal lining within the abdominal cavity and has a significant impact on patient prognosis. The success of treatments such as cytoreductive surgery (CRS) and hyperthermic intraperitoneal chemotherapy (HIPEC) depends heavily on the accurate detection of lesions through laparoscopic imaging. However, lesion detection in laparoscopic videos remains challenging due to the presence of small low-contrast lesions, large lesions that blend with organ surfaces, and visual disturbances caused by camera motion and other intraoperative factors.
This study adopts a multifaceted deep learning approach, which combines multiple strategies to address the challenges of PC lesion detection, including the evaluation of model training strategies, architectural modifications to improve the detection capability for small lesions, and the development of a preliminary lesion statistical analysis approach for laparoscopic videos. The dataset used consists of real laparoscopic images annotated by experts from the Institut de Cancérologie de l'Ouest (ICO), France. Three approaches were evaluated to improve YOLO-based lesion detection performance: (1) comparison between training from scratch and transfer learning using pre-trained weights; (2) modification of the YOLOv8 architecture through the integration of the Convolutional Block Attention Module (CBAM) and the addition of a P2 layer; and (3) development of a video analysis system based on unique frame selection using ResNet50 deep feature extraction and cosine similarity to generate statistics on detected lesions and estimated lesion areas from segmentation results.
The experimental results demonstrate that the transfer learning approach using pre-trained weights significantly outperformed training from scratch, with the pre-trained YOLOv8s achieving an mAP50 of 0.871 and an F1-Score of 0.840, while training from scratch only reached an mAP50 of 0.799 with YOLOv8l. Meanwhile, the integration of CBAM and the addition of a P2 layer to the YOLOv8 architecture improved the model's ability to detect small lesions, achieving a precision of 0.887, a recall of 0.769, an mAP50 of 0.861, and an F1-Score of 0.824. Ablation studies demonstrate that each component, namely CBAM, the P2 layer, and the CIoU loss function, contributes to the overall performance improvement. In addition, the unique frame selection approach effectively reduces frame redundancy in laparoscopic videos, enabling more efficient lesion analysis. The developed system generates statistics on detected lesions and estimated lesion areas based on the selected frames, providing an initial step toward the development of video-based lesion burden analysis systems. The results also indicate that higher cosine similarity thresholds preserve more lesion-related information, while lower thresholds provide more aggressive frame reduction at the risk of losing important lesion-containing frames.
This study contributes to the development of computer vision–based automated lesion detection and analysis systems for laparoscopic imaging and provides opportunities for the development of more comprehensive lesion analysis systems to support clinical decision-making in future oncological surgery practices.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Uncontrolled Keywords: | peritoneal carcinomatosis, laparoskopi, YOLO, pembelajaran mendalam, bobot pre-trained, transfer learning, mekanisme perhatian, CBAM, multifaset, Restnet50, cosine similarity, seleksi frame unik, peritoneal carcinomatosis, laparoscopy, YOLO, deep learning, pre-trained weights, transfer learning, attention mechanism, CBAM, multifaceted approach, ResNet50, cosine similarity, unique frame |
| Subjects: | T Technology > T Technology (General) > T57.5 Data Processing |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science) |
| Depositing User: | Naim Rochmawati |
| Date Deposited: | 29 Jul 2026 05:49 |
| Last Modified: | 29 Jul 2026 05:49 |
| URI: | http://repository.its.ac.id/id/eprint/139751 |
Actions (login required)
 |
View Item |