Multi-slice Channelized Hotelling Observer Research Articles

Strategy to implement a convolutional neural network based ideal model observer via transfer learning for multi-slice simulated breast CT images

Objective. In this work, we propose a convolutional neural network (CNN)-based multi-slice ideal model observer using transfer learning (TL-CNN) to reduce the required number of training samples. Approach. To train model observers, we generate simulated breast CT image volumes that are reconstructed using the FeldkampDavisKress algorithm with a ramp and Hanning-weighted ramp filter. The observer performance is evaluated on the background-known-statistically (BKS)/signal-known-exactly task with a spherical signal, and the BKS/signal-known-statistically task with random signal generated by the stochastic grown method. We compare the detectability of the CNN-based model observer with that of conventional linear model observers for multi-slice images (i.e. a multi-slice channelized Hotelling observer (CHO) and volumetric CHO). We also analyze the detectability of the TL-CNN for different numbers of training samples to examine its performance robustness to a limited number of training samples. To further analyze the effectiveness of transfer learning, we calculate the correlation coefficients of filter weights in the CNN-based multi-slice model observer. Main results. When using transfer learning for the CNN-based multi-slice ideal model observer, the TL-CNN provides the same performance with a 91.7% reduction in the number of training samples compared to that when transfer learning is not used. Moreover, compared to the conventional linear model observer, the proposed CNN-based multi-slice model observers achieve 45% higher detectability in the signal-known-statistically detection tasks and 13% higher detectability in the SKE detection tasks. In correlation coefficient analysis, it is observed that the filters in most of the layers are highly correlated, demonstrating the effectiveness of the transfer learning for multi-slice model observer training. Significance. Deep learning-based model observers require large numbers of training samples, and the required number of training samples increases as the dimensions of the image (i.e. the number of slices) increase. With applying transfer learning, the required number of training samples is significantly reduced without performance drop.

Channelized Hotelling observer correlation with human observers for low-contrast detection in liver CT images.

Task-based image quality procedures in CT that substitute a human observer with a model observer usually use single-slice images with uniform backgrounds from homogeneous phantoms. However, anatomical structures and inhomogeneities in organs generate noise that can affect the detection performance of human observers. The purpose of this work was to assess the impact of background type, uniform or liver, and the viewing modality, single- or multislice, on the detection performance of human and model observers. We collected abdominal CT scans from patients and homogeneous phantom scans in which we digitally inserted low-contrast signals that mimicked a liver lesion. We ran a rating experiment with the two background conditions with three signal sizes and three human observers presenting images in two reading modalities: single- and multislice. In addition, channelized Hotelling observers (CHO) for single- and multislice detection were implemented and evaluated according to their degree of correlation with the human observer performance. For human observers, there was a small but significant improvement in performance with multislice compared to the single-slice viewing mode. Our data did not reveal a significant difference between uniform and anatomical backgrounds. Model observers demonstrated a good correlation with human observers for both viewing modalities. Human observers have very similar performances in both multi- and single-slice viewing mode. It is therefore preferable to use single-slice CHO as this model is computationally more tractable than multislice CHO. However, using images from a homogeneous phantom can result in overestimating image quality as CHO performance tends to be higher in uniform than anatomical backgrounds, while human observers have similar detection performances.

Correlation between a 2D channelized Hotelling observer and human observers in a low-contrast detection task with multislice reading in CT.

Model observers have been successfully developed and used to assess the quality of static 2D CT images. However, radiologists typically read images by paging through multiple 2D slices (i.e., multislice reading). The purpose of this study was to correlate human and model observer performance in a low-contrast detection task performed using both 2D and multislice reading, and to determine if the 2D model observer still correlate well with human observer performance in multislice reading. A phantom containing 18 low-contrast spheres (6 sizes × 3 contrast levels) was scanned on a 192-slice CT scanner at five dose levels (CTDIvol = 27, 13.5, 6.8, 3.4, and 1.7 mGy), each repeated 100 times. Images were reconstructed using both filtered-backprojection (FBP) and an iterative reconstruction (IR) method (ADMIRE, Siemens). A 3D volume of interest (VOI) around each sphere was extracted and placed side-by-side with a signal-absent VOI to create a 2-alternative forced choice (2AFC) trial. Sixteen 2AFC studies were generated, each with 100 trials, to evaluate the impact of radiation dose, lesion size and contrast, and reconstruction methods on object detection. In total, 1600 trials were presented to both model and human observers. Three medical physicists acted as human observers and were allowed to page through the 3D volumes to make a decision for each 2AFC trial. The human observer performance was compared with the performance of a multislice channelized Hotelling observer (CHO_MS), which integrates multislice image data, and with the performance of previously validated CHO, which operates on static 2D images (CHO_2D). For comparison, the same 16 2AFC studies were also performed in a 2D viewing mode by the human observers and compared with the multislice viewing performance and the two CHO models. Human observer performance was well correlated with the CHO_2D performance in the 2D viewing mode [Pearson product-moment correlation coefficient R = 0.972, 95% confidence interval (CI): 0.919 to 0.990] and with the CHO_MS performance in the multislice viewing mode (R = 0.952, 95% CI: 0.865 to 0.984). The CHO_2D performance, calculated from the 2D viewing mode, also had a strong correlation with human observer performance in the multislice viewing mode (R = 0.957, 95% CI: 879 to 0.985). Human observer performance varied between the multislice and 2D modes. One reader performed better in the multislice mode (P = 0.013); whereas the other two readers showed no significant difference between the two viewing modes (P = 0.057 and P = 0.38). A 2D CHO model is highly correlated with human observer performance in detecting spherical low contrast objects in multislice viewing of CT images. This finding provides some evidence for the use of a simpler, 2D CHO to assess image quality in clinically relevant CT tasks where multislice viewing is used.

Channelized Hotelling observers for the assessment of volumetric imaging data sets

Current clinical practice is rapidly moving in the direction of volumetric imaging. For two-dimensional (2D) images, task-based medical image quality is often assessed using numerical model observers. For three-dimensional (3D) images, however, these models have been little explored so far. In this work, first, two novel designs of a multislice channelized Hotelling observer (CHO) are proposed for the task of detecting 3D signals in 3D images. The novel designs are then compared and evaluated in a simulation study with five different CHO designs: a single-slice model, three multislice models, and a volumetric model. Four different random background statistics are considered, both gaussian (noncorrelated and correlated gaussian noise) and non-gaussian (lumpy and clustered lumpy backgrounds). Overall, the results show that the volumetric model outperforms the others, while the disparity between the models decreases for greater complexity of the detection task. Among the multislice models, the second proposed CHO could most closely approach the volumetric model, whereas the first new CHO seems to be least affected by the number of training samples.