Electronic Theses and Dissertation

Malaria merupakan penyakit menular akibat parasit plasmodium yang ditularkan melalui gigitan nyamuk anopheles dan masih menjadi ancaman serius di wilayah tropis. diagnosis dini sangat penting, namun metode mikroskopis konvensional memiliki keterbatasan. deteksi malaria pada citra mikroskopis menghadapi tantangan kemiripan visual antar objek, ukuran objek kecil, serta ketidakseimbangan kelas. penelitian ini akan membandingkan kinerja dua model deep learning berbasis transfer learning, yaitu resnet50 dan inceptionv3, dalam klasifikasi parasit malaria dan prediksi letak objek menggunakan bounding box. dataset berasal dari lacuna malaria detection challenge, yang terdiri dari tiga kelas: trophozoite, wbc, dan neg. kedua model dimodifikasi menjadi arsitektur multitask dengan dua output, yaitu klasifikasi dan regresi bounding box. proses penelitian mencakup augmentasi data, resize dan normalisasi citra, serta penyesuaian hyperparameter seperti learning rate dan box loss weight. evaluasi menggunakan metrik akurasi, f1-score, intersection over union (iou), dan mean average precision (map) pada ambang iou 0,3. hasil eksperimen menunjukkan bahwa model inceptionv3 tanpa augmentasi menghasilkan performa terbaik, dengan akurasi 74,21% dan f1-score 0,76 untuk klasifikasi. pada deteksi objek, model mencapai nilai iou sebesar 0,0006 untuk kelas trophozoite, 0,0002 untuk wbc, dan 0,0000 untuk neg, serta map sebesar 0,0009 pada ambang iou 0,3. nilai iou dan map yang rendah menunjukkan bahwa model masih menghadapi tantangan dalam lokalisasi objek yang akurat. penurunan performa pada model dengan augmentasi menunjukkan bahwa transformasi spasial kurang sesuai untuk citra mikroskopis. kata kunci: malaria, deep learning, transfer learning, resnet50, inceptionv3, citra mikroskopis

Malaria is an infectious disease caused by the Plasmodium parasite, transmitted through the bites of Anopheles mosquitoes, and remains a serious threat in tropical regions. Early diagnosis is crucial; however, conventional microscopic methods have limitations. Detecting malaria in microscopic images faces challenges such as visual similarity among objects, small object sizes, and class imbalance. This study will compare the performance of two deep learning models based on transfer learning, namely ResNet50 and InceptionV3, in classifying malaria parasites and predicting object locations using bounding boxes. The dataset is sourced from the Lacuna Malaria Detection Challenge, which consists of three classes: Trophozoite, WBC, and NEG. Both models were modified into a multitask architecture with two outputs: classification and bounding box regression. The research process includes data augmentation, image resizing and normalization, as well as hyperparameter adjustments such as learning rate and box loss weight. Evaluation will use metrics such as accuracy, F1-score, Intersection over Union (IoU), and mean Average Precision (mAP) at an IoU threshold of 0.3. Experimental results show that the InceptionV3 model without augmentation achieved the best performance, with an accuracy of 74.21% and an F1-score of 0.76 for classification. In object detection, the model achieved an IoU of 0.0006 for the Trophozoite class, 0.0002 for WBC, and 0.0000 for NEG, along with an mAP of 0.0009 at an IoU threshold of 0.3. The low IoU and mAP values indicate that the model still faces challenges in accurately localizing objects. The performance drop in the model with augmentation suggests that spatial transformations are less suitable for microscopic images. Keywords: Malaria, Deep Learning, Transfer Learning, ResNet50, InceptionV3, Microscopic Images