Purpose: We investigate the impact of various deep-learning-based methods for detecting and segmenting metastases with different lesion volume sizes on 3D brain MR images.
Approach: A 2.5D U-Net and a 3D U-Net were selected. We also evaluated weak learner fusion of the prediction features generated by the 2.5D and the 3D networks. A 3D fully convolutional one-stage (FCOS) detector was selected as a representative of bounding-box regression-based detection methods. A total of 422 3D post-contrast T1-weighted scans from patients with brain metastases were used. Performances were analyzed based on lesion volume, total metastatic volume per patient, and number of lesions per patient.
Results: The performance of detection of the 2.5D and 3D U-Net methods had recall of >0.83 and precision of >0.44 for lesion volume >0.3 cm3 but deteriorated as metastasis size decreased below 0.3 cm3 to 0.58 to 0.74 in recall and 0.16 to 0.25 in precision. Compared the two U-Nets for detection capability, high precision was achieved by the 2.5D network, but high recall was achieved by the 3D network for all lesion sizes. The weak learner fusion achieved a balanced performance between the 2.5D and 3D U-Nets; particularly, it increased precision to 0.83 for lesion volumes of 0.1 to 0.3 cm3 but decreased recall to 0.59. The 3D FCOS detector did not outperform the U-Net methods in detecting either the small or large metastases presumably because of the limited data size.
Conclusions: Our study provides the performances of four deep learning methods in relationship to lesion size, total metastasis volume, and number of lesions per patient, providing insight into further development of the deep learning networks.
Market's demands of digital cameras for higher sensitivity capability under low-light conditions are remarkably
increasing nowadays. The digital camera market is now a tough race for providing higher ISO capability. In this paper,
we explore an approach for increasing maximum ISO capability of digital cameras without changing any structure of an
image sensor or CFA. Our method is directly applied to the raw Bayer pattern CFA image to avoid non-linearity
characteristics and noise amplification which are usually deteriorated after ISP (Image Signal Processor) of digital
cameras. The proposed method fuses multiple short exposed images which are noisy, but less blurred. Our approach is
designed to avoid the ghost artifact caused by hand-shaking and object motion. In order to achieve a desired ISO image
quality, both low frequency chromatic noise and fine-grain noise that usually appear in high ISO images are removed
and then we modify the different layers which are created by a two-scale non-linear decomposition of an image. Once
our approach is performed on an input Bayer pattern CFA image, the resultant Bayer image is further processed by ISP
to obtain a fully processed RGB image. The performance of our proposed approach is evaluated by comparing SNR
(Signal to Noise Ratio), MTF50 (Modulation Transfer Function), color error ∝E*ab and visual quality with reference
images whose exposure times are properly extended into a variety of target sensitivity.
In this paper, we present a new noise estimation and reduction scheme to restore images degraded by image sensor noise. Since the characteristic of noise deviates according to camera response function (CRF) and the sensitivity of image sensors, we build a noise profile by using test charts for accurate noise estimation. By using the noise profile, we develop simple and fast noise estimation scheme which can be appropriately used for digital cameras. Our noise removal method utilizes the result of the noise estimation and applies several adaptive nonlinear filters to give the best image quality against high ISO noise. Experimental results show that the proposed method yields significantly good performance for images corrupted by both synthetic sensor noise and real sensor noise.
Digital images captured from CMOS image sensors suffer Gaussian noise and impulsive noise. To efficiently reduce the
noise in Image Signal Processor (ISP), we analyze noise feature for imaging pipeline of ISP where noise reduction
algorithm is performed. The Gaussian noise reduction and impulsive noise reduction method are proposed for proper
ISP implementation in Bayer domain. The proposed method takes advantage of the analyzed noise feature to calculate
noise reduction filter coefficients. Thus, noise is adaptively reduced according to the scene environment. Since noise is
amplified and characteristic of noise varies while the image sensor signal undergoes several image processing steps, it is
better to remove noise in earlier stage on imaging pipeline of ISP. Thus, noise reduction is carried out in Bayer domain
on imaging pipeline of ISP. The method is tested on imaging pipeline of ISP and images captured from Samsung 2M
CMOS image sensor test module. The experimental results show that the proposed method removes noise while
effectively preserves edges.