A manifold learning regularization approach to enhance 3D CT image-based lung nodule classification.
- 作者列表："Ren Y","Tsai MY","Chen L","Wang J","Li S","Liu Y","Jia X","Shen C
PURPOSE:Diagnosis of lung cancer requires radiologists to review every lung nodule in CT images. Such a process can be very time-consuming, and the accuracy is affected by many factors, such as experience of radiologists and available diagnosis time. To address this problem, we proposed to develop a deep learning-based system to automatically classify benign and malignant lung nodules. METHODS:The proposed method automatically determines benignity or malignancy given the 3D CT image patch of a lung nodule to assist diagnosis process. Motivated by the fact that real structure among data is often embedded on a low-dimensional manifold, we developed a novel manifold regularized classification deep neural network (MRC-DNN) to perform classification directly based on the manifold representation of lung nodule images. The concise manifold representation revealing important data structure is expected to benefit the classification, while the manifold regularization enforces strong, but natural constraints on network training, preventing over-fitting. RESULTS:The proposed method achieves accurate manifold learning with reconstruction error of ~ 30 HU on real lung nodule CT image data. In addition, the classification accuracy on testing data is 0.90 with sensitivity of 0.81 and specificity of 0.95, which outperforms state-of-the-art deep learning methods. CONCLUSION:The proposed MRC-DNN facilitates an accurate manifold learning approach for lung nodule classification based on 3D CT images. More importantly, MRC-DNN suggests a new and effective idea of enforcing regularization for network training, possessing the potential impact to a board range of applications.
目的: 肺癌的诊断需要放射科医师检查CT图像中的每个肺结节。这样的过程可能非常耗时，并且准确性受到许多因素的影响，例如放射科医师的经验和可用的诊断时间。为了解决这个问题，我们提出开发一个基于深度学习的系统来自动分类肺结节的良恶性。 方法: 提出的方法自动确定良性或恶性给定的三维CT图像补片肺结节，以协助诊断过程。由于数据之间的真实结构经常嵌入低维流形，我们开发了一种新的流形正则化分类深度神经网络 (mrc-dnn)，直接基于肺结节图像的流形表示进行分类。揭示重要数据结构的简洁流形表示有望使分类受益，而流形正则化则对网络训练施加强大但自然的约束，防止过度拟合。 结果: 该方法在真实肺结节CT图像数据上实现了精确的流形学习，重建误差为30 HU。此外，测试数据的分类准确率为0.90，灵敏度为0.81，特异性为0.95，优于最先进的深度学习方法。 结论: 所提出的mrc-dnn有助于基于3D CT图像的肺结节分类的准确流形学习方法。更重要的是，mrc-dnn提出了一种新的有效的思想，即对网络训练实施正则化，对板的应用范围具有潜在的影响。
METHODS:OBJECTIVES:The aim was to evaluate the image quality and sensitivity to artifacts of compressed sensing (CS) acceleration technique, applied to 3D or breath-hold sequences in different clinical applications from brain to knee. METHODS:CS with an acceleration from 30 to 60% and conventional MRI sequences were performed in 10 different applications in 107 patients, leading to 120 comparisons. Readers were blinded to the technique for quantitative (contrast-to-noise ratio or functional measurements for cardiac cine) and qualitative (image quality, artifacts, diagnostic findings, and preference) image analyses. RESULTS:No statistically significant difference in image quality or artifacts was found for each sequence except for the cardiac cine CS for one of both readers and for the wrist 3D proton density (PD)-weighted CS sequence which showed less motion artifacts due to the reduced acquisition time. The contrast-to-noise ratio was lower for the elbow CS sequence but not statistically different in all other applications. Diagnostic findings were similar between conventional and CS sequence for all the comparisons except for four cases where motion artifacts corrupted either the conventional or the CS sequence. CONCLUSIONS:The evaluated CS sequences are ready to be used in clinical daily practice except for the elbow application which requires a lower acceleration. The CS factor should be tuned for each organ and sequence to obtain good image quality. It leads to 30% to 60% acceleration in the applications evaluated in this study which has a significant impact on clinical workflow. KEY POINTS:• Clinical implementation of compressed sensing (CS) reduced scan times of at least 30% with only minor penalty in image quality and no change in diagnostic findings. • The CS acceleration factor has to be tuned separately for each organ and sequence to guarantee similar image quality than conventional acquisition. • At least 30% and up to 60% acceleration is feasible in specific sequences in clinical routine.
METHODS:BACKGROUND:The main surgical techniques for spontaneous basal ganglia hemorrhage include stereotactic aspiration, endoscopic aspiration, and craniotomy. However, credible evidence is still needed to validate the effect of these techniques. OBJECTIVE:To explore the long-term outcomes of the three surgical techniques in the treatment of spontaneous basal ganglia hemorrhage. METHODS:Five hundred and sixteen patients with spontaneous basal ganglia hemorrhage who received stereotactic aspiration, endoscopic aspiration, or craniotomy were reviewed retrospectively. Six-month mortality and the modified Rankin Scale score were the primary and secondary outcomes, respectively. A multivariate logistic regression model was used to assess the effects of different surgical techniques on patient outcomes. RESULTS:For the entire cohort, the 6-month mortality in the endoscopic aspiration group was significantly lower than that in the stereotactic aspiration group (odds ratio (OR) 4.280, 95% CI 2.186 to 8.380); the 6-month mortality in the endoscopic aspiration group was lower than that in the craniotomy group, but the difference was not significant (OR=1.930, 95% CI 0.835 to 4.465). A further subgroup analysis was stratified by hematoma volume. The mortality in the endoscopic aspiration group was significantly lower than in the stereotactic aspiration group in the medium (≥40-<80 mL) (OR=2.438, 95% CI 1.101 to 5.402) and large hematoma subgroup (≥80 mL) (OR=66.532, 95% CI 6.345 to 697.675). Compared with the endoscopic aspiration group, a trend towards increased mortality was observed in the large hematoma subgroup of the craniotomy group (OR=8.721, 95% CI 0.933 to 81.551). CONCLUSION:Endoscopic aspiration can decrease the 6-month mortality of spontaneous basal ganglia hemorrhage, especially in patients with a hematoma volume ≥40 mL.
METHODS:OBJECTIVE:The primary purpose of this study was to evaluate the effectiveness of a three-dimensional (3D) software tool (smart planes) for displaying fetal brain planes, and the secondary purpose was to evaluate its accuracy in performing automatic measurements. MATERIAL AND METHODS:This prospective study included singleton fetuses with a gestational age (GA) greater than 18 weeks. Transabdominal two-dimensional ultrasound (2DUS) and 3D smart planes images were respectively used to obtain the basic planes of the fetal brain, with five parameters measured. The images, by either two-dimensional (2D) manual or 3D automatic operation, were reviewed by two experienced sonographers. The agreements between two measurements were analyzed. RESULTS:A total of 226 cases were included. The rates of successful detection by automatic display were as high as 80%. There was substantial agreement between the measurements of the biparietal diameter, head circumference and transcerebellar diameter, but poor agreement between the measurements of cisterna magna and lateral ventricle width. CONCLUSIONS:Smart Planes might be valuable for the rapid evaluation of fetal brain, because it simplifies the evaluation process. However, the technology requires improvement. In addition, this technology cannot replace the conventional manual US scans; it can only be used as an additional approach.