Date of Publication
3-2026
Document Type
Master's Thesis
Degree Name
Master of Science in Applied Physics
Subject Categories
Physics
College
College of Science
Department/Unit
Physics
Thesis Advisor
Edgar A. Vallar
Defense Panel Chair
Maria Cecilia D. Galvez
Defense Panel Member
Jowi Tsidkenu P. Cruz
Rizza Pua
Abstract (English)
Improving the accuracy of thyroid nodule diagnosis is important to avoid unnecessary treatment while still detecting cancer early. Although ultrasound is commonly used as an imaging modality, deep learning models often perform poorly when applied to images from different devices due to domain differences and data imbalance. Also, segmentation and classification are usually treated separately, even though they are closely related in clinical practice. This study proposes a domain-adaptive multi-task learning framework based on a modified U-Net architecture for joint thyroid nodule segmentation and classification. The framework integrates a Bidirectional Image Translation Network for offline domain harmonization and Fourier Domain Adaptation for low-frequency amplitude alignment during training. MSAG is incorporated into the segmentation branch to enhance spatial feature refinement. Furthermore, a geometry-aware shape–margin auxiliary loss is introduced, where morphological descriptors derived from predicted segmentation masks are coupled with malignancy prediction to encourage structural consistency between segmentation and classification outputs. The framework was evaluated on two heterogeneous public datasets, the Thyroid Nodule Classification Dataset (TNCD) and Digital Database Thyroid Image (DDTI) and a fully unseen external dataset, Thyroid Ultrasound Cine-clip (TUC). Segmentation performance was evaluated using Dice Similarity Coefficient and Intersection over Union, while classification was evaluated using Accuracy, Balanced Accuracy, Sensitivity, Specificity, F1-score, Positive Predictive Value, Negative Predictive Value, Area Under the ROC Curve (AUC), and Precision–Recall (PR-AUC) to account for domain and class imbalance. Results show that domain adaptation consistently improves minority-domain (DDTI) segmentation performance and enhances external generalization on TUC, while in 3 classification, domain adaptation strengthens threshold-independent discrimination (AUC and PR-AUC), Muti-Scale Attention Gate primarily increases Sensitivity with moderate improvements in class balance, and the Shape–Margin constraint substantially improves Specificity under malignant-skewed conditions. The proposed framework demonstrates stable and balanced performance across heterogeneous domains, achieving the highest Balanced Accuracy on the Combined and TNCD datasets while maintaining competitive discrimination on the unseen TUC dataset. Comparative benchmarking against U-Net, CMU-Net, ACS-Net, DSMA-Net, ResNet34, and VGG16 indicate competitive in-domain performance and improved cross-domain robustness. Overall, integrating data-level harmonization, low frequency-domain alignment, attention mechanism, and geometric regularization within a unified multi-task framework enhances robustness in heterogeneous thyroid ultrasound analysis. Future work will focus on exploring multi-class TI-RADS prediction, multi-center validation, acquisition-aware modeling, probability calibration, and model compression to further support clinical deployment.
Abstract Format
html
Abstract (Filipino)
Mahalagang mapabuti ang katumpakan ng diyagnosis ng bukol sa thyroid gland upang maiwasan ang hindi kailangang paggamot habang tinutukoy pa rin ang kanser. Karaniwang ginagamit ang ultrasound bilang paraan ng pagkuha ng imahe, ngunit madalas na hindi maganda ang kinalalabasan ng mga deep learning model kapag ginamit sa mga larawan mula sa iba’t ibang aparato dahil sa pagkakaiba-iba ng data at hindi pantay na dami ng datos. Bukod dito, ang segmentation at classification ay kadalasang hiwalay na ginagawa, kahit na magkaugnay sila sa aktwal na klinikal na paggamit.
Sa pag-aaral na ito, iminungkahi ang isang nakakapag-agpang sa saklaw na multi-task learning framework na nakabase sa binagong arkitektura ng modelong U-Net para sa sabay na pagse-segmentation at classification ng thyroid nodule. Gumamit rin ng modyul na tawag ay Bidirectional Image Translation Network upang gumawa ng larawang artipisyal at Fourier Domain Adaptation upang ihanay ang mababang frequency na bahagi ng imahe habang nagsasanay ang modelo. Mayroon ding MSAG na idinagdag sa sangay na segmentation upang mapabuti ang pagkuha ng mga tampok na spatial. Dagdag pa rito, isang geometry-aware na shape–margin auxiliary loss ay isinama sa sistema ng pagsasanay ng modela, kung saan ang hugis ng nodule mula sa segmentation ay ginagamit kasama ng prediksyon ng malignancy upang masigurong magkatugma ang resulta ng segmentation at classification.
Isinuri ang framework gamit ang dalawang magkaibang datos na bukas sa publiko: Thyroid Nodule Classification Dataset (TNCD) at Digital Database Thyroid Image (DDTI), pati na rin ang isang ganap na bagong external dataset na Thyroid Ultrasound Cine-clip (TUC). Ang segmentation ay sinusuri gamit ang Dice Similarity Coefficient at Intersection over Union, habang ang classification naman ay sinusuri gamit ang Accuracy, Balanced Accuracy, Sensitivity, Specificity, F1-score, Positive Predictive Value, Negative Predictive Value, AUC, at PR-AUC upang isaalang-alang ang pagkakaiba ng datos at hindi balanseng klase.
Ipinakita ng resulta na ang domain adaptation ay nakakatulong sa pagpapabuti ng segmentation lalo na sa mas kaunting data (DDTI) at pinapabuti ng performance sa bagong dataset (TUC). Sa classification naman, pinapalakas nito ang kakayahan ng modelo na makapaghiwalay ng tama ang mga klase (AUC at PR-AUC). Ang Multi-Scale Attention Gate ay nakatulong lalo na sa pagtaas ng Sensitivity, habang ang Shape–Margin constraint ay nagpaangat ng Specificity lalo na kapag mas marami ang klase ng malignant. Ipinakita ng ipinanukalang framework ang matatag at balanseng performance sa iba’t ibang datos, na may pinakamataas na Balanced Accuracy sa Combined at TNCD datos, at medyo maayos na performance sa datos na hindi ginamit sa training o ang TUC. Kumpara sa ibang modelo tulad ng U-Net, CMU-Net, ACS-Net, DSMA-Net, ResNet34, at VGG16, nagpakita ang ipinanukalang framework ng mas mahusay na performance, lalo na sa paggamit sa iba’t ibang datos.
Sa kabuuan, ang pagsasama ng pag-aayos ng datos, pag-aayon ng frequency, attention mechanism, at paggamit ng impormasyon ng hugis sa iisang multi-task framework ay nakakatulong upang maging mas matibay ang performance ng modelo sa iba’t ibang uri ng datos ng thyroid ultrasound. Sa hinaharap, tututok ang pag-aaral sa multi-class na prediksyon ng TI-RADS, pagsusuri gamit ang datos mula sa iba’t ibang ospital, mas masusing pagmomodelo, pag-aayos ng mga probability output, at pagpapagaan ng modelo upang mas maging handa ito sa paggamit sa klinika.
Abstract Format
html
Language
English
Format
Electronic
Keywords
Thyroid gland—Diseases—Diagnosis; Ultrasonic imaging
Recommended Citation
Tuico, K. A. (2026). Knowledge-augmented deep multi-task learning for nodule segmentation and classification in thyroid ultrasound images. Retrieved from https://animorepository.dlsu.edu.ph/etdm_physics/30
Upload Full Text
wf_yes
Embargo Period
4-22-2029
