Pulmonary Nodule Detection Based on ISODATA-Improved Faster RCNN and 3D-CNN with Focal Loss

Abstract

The early diagnosis of pulmonary cancer can significantly improve the survival rate of patients, where pulmonary nodules detection in computed tomography images plays an important role. In this article, we propose a novel pulmonary nodule detection system based on convolutional neural networks (CNN). Our system consists of two stages, pulmonary nodule candidate detection and false positive reduction. For candidate detection, we introduce Iterative Self-Organizing Data Analysis Techniques Algorithm (ISODATA) to Faster Region-based Convolutional Neural Network (Faster R-CNN) model. For false positive reduction, a three-dimensional convolutional neural network (3D-CNN) is employed to completely utilize the three-dimensional nature of CT images. In this network, Focal Loss is used to solve the class imbalance problem in this task. Experiments were conducted on LUNA16 dataset. The results show the preferable performance of the proposed system and the effectiveness of using ISODATA and Focal loss in pulmonary nodule detection is proved.

Pulmonary Nodule Detection Based on Faster R-CNN With Adaptive Anchor Box

“…For instance, the second-highest sensitivity is 96.8% (a 0.7% difference with ours) by Wang et al, 2019 [24], but at 27 times more false positives. For another comparison, ISODATA (a K-meansbased clustering method) for anchor box sizes, proposed in [8], achieved 91.4% (a 6.1% difference with ours) at 3.19 FPs/s (1.45 times more false positives than ours). These results imply that the adaptive anchors generated by our Meanshift based method are more effective for nodule candidate detection than other manually designed anchors and other clustering methods, such as K-means or ISODATA.…”

“…The authors obtained a sensitivity of 90% with 15 false positives per scan and a CPM score of 90.3% on LUNA16 dataset. In a more general approach, Tong et al, 2020 [8] proposed Iterative Self-Organizing Data Analysis Techniques Algorithm (ISODATA) for analyzing anchor box to Faster R-CNN in order to further improve candidate detection performance. This is an extension of K-means clustering to automatically merge and split clusters during its iteration process, making it generally more adaptive than the original K-means method.…”

“…[23], [24], [33]), or K-means-based anchor box clustering (e.g. [8]) approaches. To the best of our knowledge, no work in the literature has investigated using Mean-shift clustering method to perform anchor box adaptation.…”

“…Traditional CAD systems usually generate nodule candidates using hand-crafted features of nodules such as morphological features [4] or intensity thresholding [5], and perform VOLUME 4, 2016 false positive reduction using conventional classifiers such as K-nearest neighbor [6] or Support Vector Machine [7]. However, these methods rarely achieve satisfactory results due to the difficulty in distinguishing nodules from other lung structures, like blood vessels or lung tissues [8], mainly due to the great variability of nodule features.…”

“…Nowadays, deep learning-based object detection methods have become dominant in CAD systems for lung nodule detection. Among them, methods based on Faster Regionbased Convolutional Neural Network (Faster R-CNN) [9], a two-stage object detection method, have gained enormous research attention [8] because of their leading performance in lung nodule candidate detection [10]. A two-stage method firstly proposes as many Regions-of-Interest (RoIs) as possible, then secondly classifies each region and refines its bounding box.…”