Two-stage lung nodule detection framework using enhanced UNet and convolutional LSTM networks in CT images

Lung cancer is one of the leading causes of cancer deaths globally, and lung nodules are the primary indicators that aid in early detection. The computer-aided detection (CADe) system acts as a second reader, reducing the variability in lung cancer risk assessment across physicians. This work aims to improve the performance of CADe systems by developing high sensitive and resilient detection networks using deep learning. This paper proposes a novel CADe framework to detect nodules from CT scans using an enhanced UNet in conjunction with a pyramid dilated convolutional long short term memory (PD-CLSTM) network. The proposed CADe system works in two stages: nodule detection and false nodule elimination. In the first stage, a modified UNet-based model, Atrous UNet+, is proposed to detect nodule candidates from axial slices using dilation and ensemble mechanisms. Dilated convolution is a powerful technique for dense prediction by incorporating larger context information without increasing the computation load. Ensemble skip connections fuse multilevel semantic features that help detect nodules of diverse sizes. In the second stage, The pyramid dilated convolutional LSTM network is proposed to identify true nodules using inter-slice and intra-slice spatial features of 3D nodule patches. In this work, a novel idea of applying convolution long short-term memory (ConvLSTM) is attempted to categorize true nodules from false nodules and help to eliminate false nodules. Experimental results on the LUNA16 dataset show that our proposed CADe system achieves the best average sensitivity of 0.930 at seven predefined FPRs: 1/8, 1/4, 1/2, 1, 2, 4, and 8 FPs per scan. Also, the proposed CADe system detects small nodules in the range of 5-9 mm with a sensitivity of 0.92 and other nodules (>10 mm) with a sensitivity of 0.93, resulting in a high detection rate in recognizing nodules of diverse sizes.