While diagnostic imaging is the best tool for early detection of bone metastases, the imaging modalities currently available — CT, MRI, bone scintigraphy, PET and PET/CT — often present challenges, according to presenters of a Sunday session.
"Although CT enables imaging bone density and morphological structures without the effects caused by the superimpositions of bones and organs, its low sensitivity (74%) and specificity (56%) means it has limited use as a bone metastases screening test," said presenter Saori Koshino, MD, a radiologist at the University of Tokyo's Graduate School of Medicine.
Nevertheless PET/CT does allow users to more accurately identify the location of a lesion with superior image quality. In fact, its sensitivity and specificity are 98 and 56% respectively. However, PET/CT has the significant disadvantage of being costly and of limited availability.
Deep Learning May Lead to Gold Standard
In recent years, computer aided diagnosis (CAD) has started to adopt deep learning (DL) approaches. "Generally, data annotation has been one of the most relevant aspects of the whole development of CAD systems using DL," Dr. Koshino said. "Yet despite these advancements, due to the large volume of available data, there is no gold standard for generating annotations in the radiological field."
To achieve such a standard, Dr. Koshino has developed an algorithm using DL for the early detection of bone metastases from CT data. Her research also explored the feasibility of automatically generating ground-truth data for training the proposed DL algorithm using PET/CT datasets.
The end goal was to produce an artificial intelligence (AI) pipeline system based on 18F‑fluorodeoxyglucose (FDG) PET/CT for the automatic annotation and detection of bone metastases from CT data. "We hypothesized that the proposed detection method would achieve high sensitivity and that the automated annotation method would help reduce costs and mitigate the time-consuming task of data annotation," Dr. Koshino said.
The study retrospectively evaluated 201 whole-body PET/CT examinations with bone metastases. First, an automated annotation tool for bone metastases was created to extract lesions from PET data. A binary bone mask was obtained from CT data and multiplied by the rescaled PET volume. A convolutional neural network (CNN), ResNet-50, was used to discard high FDG uptake regions outside the bones.
Each bone metastasis candidate was then converted into a slice-wise bounding box. Afterward, a radiologist labeled automatically extracted lesions with osteoblastic metastasis, osteolytic metastasis, intertrabecular metastasis or other forms of metastases.
Using the automatic annotation tool, 1,377 lesions were detected, consisting of 403 osteoblastic metastases, 381 osteolytic metastases, 30 intertrabecular metastases, and 563 other forms of metastases, including normal regions and inflammation. The lesion-wise sensitivity of the Mask R-CNN model on the validation dataset was 77.8% (21/27) for osteoblastic metastases and 74.6% (44/59) for osteolytic metastases, with a false positive per image of 0.71.
According to Dr. Koshino, this research is an important step toward leveraging AI for better detection and increased accuracy of bone metastases using CT data. "This proposed AI pipeline system is well-positioned to alleviate the burden of generating annotations and prevent oversights by radiologists when detecting bone metastases," she added.