Page 226 - Hospital Authority Convention 2018

Service Enhancement Presentations
HOSPITAL AUTHORITY CONVENTION 2018

F8.5 Young HA Investigators Session 14:30 Room 421

Pilot Study on Machine Learning for Bone Scan Classification
1
2
1
2
Fang XHB , Ho WY , Ma WHV , Lam KCA , Lam WMW 1
1 2
Department of Radiology, Nuclear Medicine Unit, Department of Radiology, Queen Mary Hospital, Hong Kong
Introduction
The performance of artificial intelligence in image recognition tasks has improved drastically over the past decade. This is
mainly due to one form of machine learning algorithm called convolutional neural network (CNN). In this pilot study, we test
the performance of a simple CNN on bone scan classification.
Objective
We test the performance of a simple CNN trained with a small dataset on the binary classification of bone scan images on
whether bone metastasis is present or not.

Methodology
We constructed a simple CNN which has the structure of 4 convolutional layers followed by 2 fully connected layers. The
input layer is a 512x512 single channel image. The final output layer is a binary classification of whether bone metastasis
is present or not. The design of the network is an arbitrary balance between complexity and easiness on memory and
processing requirements. Training was performed on a desktop PC with two GeForce GTX 780Ti graphics cards. We used
TensorFlow as the deep learning framework. Our dataset consisted of 106 labeled anonymised DICOM images. Each image
comprised an anterior whole body scan on the left side and a posterior whole body scan on the right side. Labelling was
performed by a nuclear medicine specialist who has over 29 years of experience in bone scan interpretation. 48 images were
labelled as without bone metastasis and 58 of them as with bone metastasis. One-third of the images were randomly set
aside for testing while the remaining two-thirds of images were used for training and validation of the CNN model. Images
used for training were augmented with random rotation, translation, zooming and occlusion.
Results
The CNN achieved 100% accuracy on the training images and >90% accuracy on the test images.

Tuesday, 8 May 2018

224