Hounsfield Unit Variations-Based Liver Lesions Detection and Classification Using Deep Learning.

This paper discusses about a method adopted to develop a computer-aided diagnostic system to achieve automatic detection and classification of liver lesions. The procedure followed consists of first segmenting the CT scan image so as to accurately extract out the lesion region alone from the rest of the abdominal details. This Region Of Interest(ROI) is now used up for extracting out first order and second order statistical feature values, which aids in the correct classification of lesions. The lesions can be classified into five types: normal liver, cysts, abscesses, benign growth (hemangioma, focal nodular hyperplasia, hepatocellular adenoma etc) and malignant growth (Hepatocellular Carcinoma, metastases etc), and this paper discusses a robust method for correctly identifying and classifying these lesions of the liver.

With the increasing daily workload of physicians, computer-aided diagnosis (CAD) systems based on deep learning play an increasingly important role in pattern recognition of diagnostic medical images. In this paper, we propose a framework based on hierarchical convolutional neural networks (CNNs) for automatic detection and classification of focal liver lesions (FLLs) in multi-phasic computed tomography (CT). A total of 616 nodules, composed of three types of malignant lesions (hepatocellular carcinoma, intrahepatic cholangiocarcinoma, and metastasis) and benign lesions (hemangioma, focal nodular hyperplasia, and cyst), were randomly divided into training and test sets at an approximate ratio of 3:1. To evaluate the performance of our model, other commonly adopted CNN models and two physicians were included for comparison. Our model achieved the best results to detect FLLs, with an average test precision of 82.8%, recall of 93.4%, and F1-score of 87.8%. Our model initially classified FLLs into malignant and benign and then classified them into more detailed classes. For the binary and six-class classification, our model achieved average accuracy results of 82.5 and73.4%, respectively, which were better than the other three classification neural networks. Interestingly, the classification performance of the model was placed between a junior physician and a senior physician. Overall, this preliminary study demonstrates that our proposed multi-modality and multi-scale CNN structure can locate and classify FLLs accurately in a limited dataset, and would help inexperienced physicians to reach a diagnosis in clinical practice.

To evaluate diffusion-weighted MR imaging (DWI), gadoxetic acid-enhanced MR imaging and the combination of both methods in the detection, classification, and characterization of focal liver lesions (FLL). A total of 119 FLL (28 HCCs, 39 metastases, 15 FNHs, 11 adenomas, 13 hemangiomas, 13 cysts) were retrospectively analyzed in 36 patients. In those patients MR imaging of the liver comprising respiratory-triggered DWI (b values of 50, 300, and 600 s/mm(2)) and gadoxetic acid-enhanced MR imaging including image acquisition in the hepatocyte-selective phase (20 min post injection) had been performed. Three image sets were assigned and compared: DWI only (set A), gadoxetic acid-enhanced MR imaging only (set B), and both modalities in combination (set C). Two readers independently interpreted the images in random order. For each reader and image set, the area under the receiver operating characteristic curve (Az) and sensitivity in the detection of FLL was determined as well as the accuracy in the classification and characterization of FLL. There was no significant difference between the three image sets in the detection of FLL with regards to Az. However, when only lesions with a diameter of 10 mm or less were analyzed, the Az values of set C were significantly higher than those of sets A and B for both readers. For classifying and characterizing FLL both set B and C were significantly superior to set A. Adding DWI to gadoxetic acid-enhanced MR imaging significantly increases the accuracy in the detection of small FLL.

ObjectiveT2 mapping allows for the classification of focal liver lesions, differentiating malignancies from the most common benign liver lesions, hemangiomas, and bile duct hamartomas (BDH). Partial volume (PV) due to the presence of liver and lesion within the same voxel confounds the classification of small lesions. Our objective is to develop a robust two-component T2 estimation technique (SEPG2-SP) to enable accurate T2 estimation in the presence of PV.Materials and methodsT2 estimation accuracy was evaluated using computer simulations, physical phantom data, and in vivo in 27 subjects with focal liver lesions (16 males, 62.4 ± 14.3 years old; 11 females, 66.8 ± 5.8 years old) imaged at 1.5 T with a radial turbo spin-echo (RADTSE) technique. The SEPG2-SP model was compared to a single-component model, which does not account for PV. The area under the receiver operator characteristic curve (AUROC) was used to analyze lesion classification.ResultsPhantom data showed that the SEPG2-SP model had a T2 estimation error of 2–9% while the single component model had a larger error of 9–23%. Analysis of in vivo data from 68 focal liver lesions (33 malignancies, 7 hemangiomas, and 28 BDH) showed that the SEPG2-SP model classified all lesions correctly (AUROC = 1), regardless of their size. On the other hand, with the single-component model, there was overlap between malignancies and benign lesions driven by misclassification of hemangiomas as malignancies (AUROC = 0.84).ConclusionsThe two-component T2 model improved the characterization of focal liver lesions affected by PV, yielding complete separation of malignancies from the most common benign liver lesions.

Tensor-based sparse representations of multi-phase medical images for classification of focal liver lesions

Computer-aided Classification of Liver Lesions from CT Images Based on Multiple ROI

Perfusion CT of the liver typically involves scanning the liver at least 20 times, resulting in a large radiation dose. We developed and validated a simplified model of tumor blood supply that can be applied to standard triphasic scans and evaluated whether this can be used to distinguish benign and malignant liver lesions. Triphasic CTs of 46 malignant and 32 benign liver lesions were analyzed. For each phase, regions of interest were drawn in the arterially enhancing portion of each lesion, as well as the background liver, aorta, and portal vein. Hepatic artery and portal vein blood supply coefficients for each lesion were then calculated by expressing the enhancement curve of the lesion as a linear combination of the enhancement curves of the aorta and portal vein. Hepatocellular carcinoma (HCC) and hypervascular metastases, on average, both had increased hepatic artery coefficients compared to the background liver. Compared to HCC, benign lesions, on average, had either a greater hepatic artery coefficient (hemangioma) or a greater portal vein coefficient (focal nodular hyperplasia or transient hepatic attenuation difference). Hypervascularity with washout is a key diagnostic criterion for HCC, but it had a sensitivity of 72% and specificity of 81% for diagnosing malignancy in our diverse set of liver lesions. The sensitivity for malignancy was increased to 89% by including enhancing lesions that were hypodense on all phases. The specificity for malignancy was increased to 97% (p = 0.039) by also examining hepatic artery and portal vein blood supply coefficients, while maintaining a sensitivity of 76%.

Liver cancer has a high mortality rate and is a serious threat to human life. The study of automated methods for analyzing liver cancer is very helpful to doctors in making a diagnosis. The existing methods tend to ignore the information correlation between multiple modalities of magnetic resonance imaging and do not design networks for multiple modalities and liver lesions. These methods are deficient in liver lesion classification and prediction performance, limiting development of the field. Therefore, we consider the information correlation between the multimodalities and design a multidimensional dual encoding network that can make full use of the information between the eight modalities to improve the classification and the prediction performance of liver lesions. It consists of a multidimensional information extraction, a dual encoder, and a classification structure. Firstly, a method for the application of multimodal data is designed, and the multidimensional information extraction module is used to extract two-dimensional (2D) and three-dimensional (3D) information from all modalities. Then, the dual encoder is used to improve feature extraction and pass multi-scale information to the classification structure. Finally, two differently connected networks were used to train the model for joint prediction, improving the final results. In this paper, a multiphase magnetic resonance imaging dataset containing 498 images was used for the experiments. The method was validated by ablation studies and comparisons with state-of-the-art (SOTA) methods, achieving balanced F1 scores, Cohen_Kappa, accuracy, and area under curve of 0.781, 0.731, 0.779, and 0.944, respectively.

The objective of our study was to evaluate a new 3D fast spoiled gradient-recalled echo (FSPGR) sequence referred to as modified liver acceleration volume acquisition (LAVA) for high-resolution gadolinium-enhanced dual arterial phase liver MRI and to determine the effect of this technique on the timing of the contrast bolus and lesion detection. Gadolinium-enhanced dual arterial phase liver MRI was performed in 109 patients using a modified LAVA sequence that supports adaptive 2D centric view ordering, efficient 2D autocalibrated acceleration, and partial-Fourier to achieve faster scan times while maintaining the same slice thickness, resolution, and coverage as single-phase imaging. After a fixed 20-second scan delay, a modified LAVA acquisition required a single 24- to 26-second breath-hold for two arterial phases with 56-60 slices per pass. Images were reviewed for timing relative to liver enhancement, lesion conspicuity, and lesion detection. Liver lesion depiction was evaluated qualitatively and quantitatively. A control group of 109 patients underwent imaging using a single arterial phase 3D FSPGR sequence, which was also performed with a fixed 20-second scan delay. The single arterial phase images produced optimal timing in the middle or late arterial phase in 79 (72%) of the 109 control group patients compared with 99 (91%) of the 109 study group patients who underwent imaging using a dynamic modified LAVA dual arterial phase sequence. For the modified LAVA sequence, the first-pass images were obtained during the mid arterial phase in 34 patients (31%). The second-pass images were obtained during the mid arterial phase in 51 patients (47%) and late arterial phase in 26 patients (24%). Sixty-two patients had liver lesions showing greater conspicuity--on the first phase in 17 patients (27%) and second phase in 45 patients (73%). Hypovascular lesions were more conspicuous on second-phase images in 24 (86%) of 28 patients. Hypervascular lesions were more conspicuous on first-phase images in 13 patients (38%) and on second-phase images in 21 (62%) of 34 patients. The first-phase images detected 165 and 155 liver lesions, respectively, for two observers compared with 233 and 224 lesions on the second-phase images, whereas the combined dual arterial phase images detected 256 and 248 hepatic lesions. High-resolution dual arterial phase 3D FSPGR MRI improves the timing of the arterial phase of liver enhancement and provides additional information for liver lesion detection.

Objective To assess gadoxetic acid (Gd-EOB-DTPA) and diffusion-weighted imaging (DWI) value in classification (benign vs. malignant) and characterization of solid focal liver lesions (SFLLs) and impact on confidence in diagnosis. Methods A total of 195 lesions (46 hepatocellular carcinomas [HCCs], 45 metastases, 32 adenomas, 37 focal nodular hyperplasias [FNHs] and 35 hemangiomas) were retrospectively evaluated in 93 patients. Three imaging datasets were compared: DWI/ apparent diffusion coefficient (ADC) (set A), Gd-EOB-DTPA (set B) and combination of both (set C). Two radiologists (R) independently classified (on a five-point ordinal scale) and characterized each lesion. The accuracy in classification and characterization was compared, and the diagnostic confidence was assessed. Results The classification accuracy on set A, B and C was 86.2%, 91.3% and 91.8% (R1), and 84.6%, 91.8% and 93.3% (R2); and characterization accuracy was 67.2%, 88.2% and 87.7% (R1), and 60.5%, 88.2% and 85.6% (R2). Classification by reader 1 showed no significant difference between set A and B (p=.09). For both readers, there was a significant difference between set A and C in both classification and characterization (all p < .05), but no significant difference between set B and C in neither classification nor characterization. No significant difference between the three datasets in classification and characterization of hemangiomas (all p > .05). The diagnostic confidence of the readers has increased progressively from set A to Set C (all p < .01). Conclusions DWI may suggest benignity or malignancy of solid liver lesions, while Gd-EOB-DTPA-enhanced imaging remains superior in lesions characterization and the combination of both increases the diagnostic confidence. DWI is very helpful in the diagnosis of hepatic hemangiomas.

Classification of liver lesions in CT images based on LivlesioNet, modified Multi-Scale CNN with bridge Scale method

In this paper, the liver lesions classification system for CT images use deep learning (CNN)model withimproved accuracy has proposed. The sequential model of CNN architecture with I/P convolution layer, Hidden&#x0D; convolution layer, and O/P convolution layer for CT images have been used to classify liver lesions. TensorFlow-2.0 is used to make an image with varying image qualities.The proposed network is used for CT images with an&#x0D; image size of 65×65, 60×60, 50×50 for which liver lesions classification accuracy of 99%,97%,95% respectivelyare achieved. The regularization technique used in proposed N/W has helped to improve the accuracy andminimization over fitting problem. The classification accuracy improvement has justified by comparing theproposed research work with other researcher's work.

The lack of training samples is one of the main factors affecting the development of deep learning methods. Deep learning models often fail to learn useful features and have serious over-fitting problems when lacking of training data. In this work, we exploit two popular unsupervised learning techniques: adversarial learning and self-supervised learning, which is aimed at mine more useful representations and relieve over-fitting problems. Our training scheme is mainly divided into three steps. Firstly, we train a self-supervision network with unsupervised learning to extract obvious features from our liver lesion samples and these features will be transferred to next step. Secondly, we use the final output feature map generated by self-supervision network to train a discriminator by adversarial learning. Finally, the backbone network is trained under the constraint of discriminator and classifier. Our main idea is to train a discriminator with adversarial learning and self-supervised learning. Then, we use the discriminator to constrain the backbone network, which is aimed to reduce the backbone network solution search space. In particular, Different from generating data with GAN, we use GAN to feature adversarial learning for feature augmentation. Our experiments on liver lesion classification in CT show an average accuracy as 92.51% compared with the baseline training scheme, which demonstrates our proposed method can mime useful features and relieve over-fitting problem. It can assist physicians in the early detection and treatment of liver lesions.KeywordsSelf-supervised learningAdversarial learningLiver lesions

Contrast-enhanced computed tomography (CT) is essential for tumor assessment, but the detection of low-contrast liver lesions remains challenging. Reducing the radiation dose increases image noise, compromising image quality and diagnostic accuracy. Iterative reconstruction (IR) algorithms can reduce noise; however, they can also alter image texture and limit lesion detection. Deep-learning image reconstruction (DLIR) represents a promising alternative, but its efficacy in ultra-low-dose (ULD) hepatic CT for detecting small, low-contrast lesions remains underexplored. Thus, this study aimed to evaluate a novel real-time DLIR algorithm in ULD hepatic CT, focusing on image quality and lesion detection. In total, 65 patients with hepatic lesions underwent both standard-dose and ULD abdominal CT scans during the portal venous phase. The standard-dose protocol (group A) used 120 kV with a signal-to-noise ratio (SNR) of 1.0, and the images were reconstructed using 50% IR. The ULD protocol (group B) used 120 kV with an SNR of 0.5, and the images were reconstructed using 50% IR and 50% DLIR (groups B1 and B2, respectively). The quantitative and qualitative image quality parameters were assessed. The lesion detection rates were evaluated by lesion type and size using the metrics of detection rate, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Group B showed a 73.3% reduction in the radiation dose compared to group A (1.5±0.8 vs. 6.9±2.0 mSv, P<0.001). Image noise differed significantly across the protocols: group B1 had the highest noise [10.05±2.94 Hounsfield units (HU)], followed by groups A (8.29±2.82 HU) and B2 (8.04±2.71 HU; all pairwise P<0.05 except group A vs. group B2: P=0.625). The CT values and contrast-to-noise ratios (CNRs) were comparable between groups B2 and A (all P>0.05), while group B2 had a 29.9-42.2% higher CNR than group B1 (all P<0.001). The qualitative assessments confirmed that the image quality and diagnostic acceptability of groups B2 (100%) and A (all P>0.05) were comparable, while the images of group B1 were diagnostically unacceptable (all scores <3). Overall, lesion detection was comparable in groups B2 (90.5%, 133/147) and A (98.0%, 144/147; P>0.05). However, group B2 had a significantly lower detection rate for small lesions (<0.5 cm: 77.8%, 42/54) compared to group A (P<0.05), but outperformed group B1 (57.4%, 31/54; P<0.05). Group B2 also had a significantly improved lesion detection rate and sensitivity for low-contrast lesions (87.2%, 95/109) compared to group B1 (75.2%, 82/109; P<0.05). The novel DLIR algorithm achieved a reconstruction speed of 60 images per second (ips), which was significantly faster than that of other DLIR approaches, while maintaining comparable performance to the IR algorithm. The combination of tube current reduction with a novel real-time DLIR algorithm enabled ULD abdominal CT to achieve a 73.3% reduction in the radiation dose while maintaining image quality and diagnostic performance for detecting hepatic lesions larger than 0.5 cm.

4071 Background: In patients with colorectal cancer detection and differentiation of liver metastases is important for further treatment decissions. Recently diffusion weighted MR imaging (DWI) was introduced as an new technique for liver tumor detection and characterzation. DWI is sensitive to molecular diffusion which is the random thermal motion of molecules (Brownian motion). MR imaging is the only method by which molecular diffusion processes can be evaluated in vivo. New techniques, such as breath-hold imaging or respiratory triggered single shot echo-planar-imaging (EPI) allow to extent DWI to abdominal organs, especially the liver. First results show that DWI can be used for the reliable detection of small subcentimeter liver lesions. We evaluated this new imaging method with regard to liver lesion detection and lesion characterization in patients with colorectal cancer. Methods: 115 consecutive patients with colorectal cancer and suspected liver metastases were evaluated with multi slice-CT and MRI at 1.5 Tesla using T1- and T2-weighted pulse sequences as well as DWI with respiratory triggered single shot echo-planar-imaging (EPI). Images were analyzed by two experienced radiologist blinded to the clinical results. Results for lesion detection for CT and DWI-MRI were compared using ROC analysis. Furthermore, in a subset of 60 patients diffusion coefficients (ADCs) were calculated for the differentiation of liver lesions (benign vs malignant). Results: For lesion detection ROC analysis showed superiority (p< 0.05) of DWI as compared to MS-CT. Especially metastases with diameters of smaller 1 cm were better detected with DWI-MRI as compared to MS-CT (p < 0.02). Results for differentiation show mean ADC-values of 1.24 for normal liver while malignant tumors have ADCs of 1.04–1.22 and benign lesions such as hemangiomas or cysts of 1.9 to 3.0, respectively. Furthermore, therapeutic changes were made in 20 % of patients based on DWI. Conclusions: MR imaging using DWI is more sensitive than multi- detector row CT in the detection and characterization of liver metatstases from colorectal cancer. Thus, it can be used as a new method for liver lesion detection and characterization as well as for optimal treatment planning. No significant financial relationships to disclose.