Assessment of Osseous Metastasis From Bone Scintigrams Using Deep Learning

Date of Award

7-1-2022

Document Type

Thesis

Degree Name

Master of Science in Data Science

First Advisor

Patricia Angela R. Abu, PhD

Abstract

The bone is a frequent site of distant metastasis for many cancers. The presence of osseous metastasis represents an advanced stage of malignancy with a median survival of a few months and with limited appropriate therapies. The consequent structural bone destruction leads to considerable morbidity, including untreatable pain, fractures, functional impairment which impact on the patient’s quality of life. Hence, it is important to make an early diagnosis of osseous metastasis to provide an accurate patient’s treatment plan with the goal of improving overall survival rates and/or quality of life. The aim of this study is to develop a deep learning model using a convolutional neural network to assess the presence of osseous metastasis from bone scintigram dataset of a local medical institution. The creation of the network architecture was made via an exploratory process combined with bibliographic search. Several experiments were made to determine batch size of the network with the highest classification performance . The base model was also compared with the pre-trained architecture used in medical image classification: (1) VGG16, (2) ResNet50, (3) DenseNet121, and (4) InceptionV3. Our findings on the batch size for all models (base and pre-trained) suggest the use of smaller batch size as larger batch sizes did not show any increase in the accuracy rates to suggest improved classification performance.

For the assessment of osseous metastasis, our results showed the best base model and the best pre-trained architecture having very good and similar accuracy rates, very good to excellent specificity and good to excellent precision rates. All models except InceptionV3 had poor to fair sensitivity and fair F1-scores while InceptionV3 had good sensitivity and F1-scores. On the basis of Matthews correlation coefficient, ResNet50 outperformed the rest of the models as to classification performance. Since all models had almost similar metric values, any of the models can be used as a decision support tool for nuclear medicine physicians in their clinical practice. In combination with the known excellent sensitivity of bone scan as an imaging modality to assess for osseous metastasis, the use of these models having very good to excellent specificity, good to excellent precision and very good accuracy rates indicates the clinical utility of these models leading to an enhanced diagnostic accuracy of bone scintigrams. Lastly, our results also showed that geometric augmentation techniques led to an improvement in the classification performance demonstrating its usefulness.

Link to Full Text

Share

COinS