Electronic Theses and Dissertation

Kardiomegali merupakan pembesaran jantung yang dapat terdeteksi melalui citra Xray dada dengan menghitung nilai Cardiothoracic Ratio (CTR). CTR adalah rasio lebar jantung terhadap toraks, nilai CTR yang lebih besar dari 0,5 menunjukkan adanya indikasi kardiomegali. Deteksi manual menggunakan metode konvensional memerlukan keahlian khusus dan sering kali bersifat subjektif. Penelitian ini bertujuan untuk menerapkan metode deep learning dalam segmentasi semantik citra X-ray dada guna memperoleh nilai CTR secara otomatis menggunakan bounding box. Penelitian ini akan membandingkan kinerja dari tiga arsitektur model deep learning, yaitu FCDenseNet, SegNet, dan UNet dalam segmentasi organ jantung dan paru-paru sebagai representasi toraks. Proses penelitian mencakup pengumpulan dan preprocessing data, melatih dan menguji model, serta hyperparameter tuning batch size, learning rate, dan optimizer untuk memperoleh model dengan performa paling optimal. Evaluasi untuk hasil segmentasi model dilakukan menggunakan metrik DiceSimilarity Coefficient (DSC) dan Intersection over Union (IoU), sedangkan evaluasi nilai CTR dilakukan menggunakan sensitivity dari klasifikasi hasil segmentasi menjadikelas kardiomegali dan non kardiomegali dengan ambang batas 0,5. Hasil penelitian menunjukkan bahwa model UNet menghasilkan performa segmentasi paling baik dengan rata-rata DSC sebesar 96% dan IoU sebesar 92,36%, diikuti oleh FCDenseNet dengan rata-rata DSC 95,91% dan IoU 92,20%, serta SegNet dengan rata-rata DSC 95,82% dan IoU 92,02%. Sementara itu, pada klasifikasi kardiomegali, model UNet juga mencatat sensitivity tertinggi sebesar 95,56%, sedangkan FCDenseNet danSegNet sama-sama memperoleh sensitivitas sebesar 95%. Temuan ini mengindikasikan bahwa pendekatan deep learning yang dikembangkan efektif dalamsegmentasi citra X-ray dada untuk mendukung deteksi otomatis kardiomegali melalui pengukuran CTR.

Cardiomegaly is an enlargement of the heart that can be detected through chest X-ray images by calculating the Cardiothoracic Ratio (CTR) value. CTR is the ratio of heart width to thorax width, and a CTR value greater than 0,5 indicates cardiomegaly. Manual detection using conventional methods requires special expertise and is often subjective. This study aims to apply deep learning methods in semantic segmentation of chest X-ray images to automatically obtain CTR values using bounding boxes. This study will compare the performance of three deep learning model architectures, namely FCDenseNet, SegNet, and UNet, in segmenting the heart and lung organs as representations of the thorax. The research process includes data collection and preprocessing, model training and testing, and hyperparameter tuning of batch size, learning rate, and optimizer to obtain the model with the most optimal performance. The evaluation of the segmentation results of the model was performed using the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics, while the evaluation of the CTR value was performed using the sensitivity of the classification of the segmentation results into cardiomegaly and non cardiomegaly classes with a threshold of 0,5. The results show that the UNet model produces the best segmentation performance with an average DSC of 96% and IoU of 92.36%, followed by FCDenseNet with an average DSC of 95.91% and IoU of 92.20%, and SegNet with an average DSC of 95.82% and IoU of 92.02%. Meanwhile, in cardiomegaly classification, the UNet model also recorded the highest sensitivity of 95.56%, while FCDenseNet and SegNet both achieved a sensitivity of 95%. These findings indicate that the developed deep learning approach is effective in segmenting chest X-ray images to support automatic detection of cardiomegaly through CTR measurement.