Transfer learning (TL) with convolutional neural networks aims to improve performances on a new task by leveraging the knowledge of similar tasks learned in advance. It has made a major contribution ...to medical image analysis as it overcomes the data scarcity problem as well as it saves time and hardware resources. However, transfer learning has been arbitrarily configured in the majority of studies. This review paper attempts to provide guidance for selecting a model and TL approaches for the medical image classification task.
425 peer-reviewed articles were retrieved from two databases, PubMed and Web of Science, published in English, up until December 31, 2020. Articles were assessed by two independent reviewers, with the aid of a third reviewer in the case of discrepancies. We followed the PRISMA guidelines for the paper selection and 121 studies were regarded as eligible for the scope of this review. We investigated articles focused on selecting backbone models and TL approaches including feature extractor, feature extractor hybrid, fine-tuning and fine-tuning from scratch.
The majority of studies (n = 57) empirically evaluated multiple models followed by deep models (n = 33) and shallow (n = 24) models. Inception, one of the deep models, was the most employed in literature (n = 26). With respect to the TL, the majority of studies (n = 46) empirically benchmarked multiple approaches to identify the optimal configuration. The rest of the studies applied only a single approach for which feature extractor (n = 38) and fine-tuning from scratch (n = 27) were the two most favored approaches. Only a few studies applied feature extractor hybrid (n = 7) and fine-tuning (n = 3) with pretrained models.
The investigated studies demonstrated the efficacy of transfer learning despite the data scarcity. We encourage data scientists and practitioners to use deep models (e.g. ResNet or Inception) as feature extractors, which can save computational costs and time without degrading the predictive power.
IntroductionSeveral mathematic models exist that purport to predict the relationship between branching vessels. In the context of diffuse atheroma, it can be difficult to appreciate if the left main ...stem is abnormally diffusely narrow relative to the other vessels. Accurately determining the normal size relationship is useful to define if disease is present. The purpose of this study is to assess the reliability of different mathematical models at predicting the size relationship between left main stem, left anterior descending (LAD) and circumflex arteries using CT coronary angiography.Methods60 consecutive CTCA examinations showing no atheroma performed as a part of a clinical service formed the study cohort. Retrospective analysis was undertaken to independently assess the diameter of the left main stem, LAD and circumflex vessels using two experienced observers. Three mathematic models were assessed to determine the accuracy of predicting left main diameter using the LAD and CX dimensions; Murray, Huo-Kassab (HK) and Finet.ResultsBased on intraclass correlation values the interobserver agreement for diameter measurement is good (0.9). Murray’s formula generated a result that was the closest to the actual diameter of the three estimates in 86.0% of cases (49 of 57). Murray was the best predictor, being significantly better than both HK (p<0.001) and Finet (p<0.001). HK was significantly better than Finet (p=0.011). The disparity in size of the two daughter vessels does not affect the ability of each equation to predict LMS size.ConclusionOur results indicate that Murray’s law is the best predictor of left main stem diameter based on the size of the LAD and circumflex coronary arteries.
Convolutional neural networks (CNNs) are deep learning network architectures that have pushed forward the state-of-the-art in a range of computer vision applications and are increasingly popular in ...medical image analysis. However, training of CNNs is time-consuming and challenging. In medical image analysis tasks, the majority of training examples are easy to classify and therefore contribute little to the CNN learning process. In this paper, we propose a method to improve and speed-up the CNN training for medical image analysis tasks by dynamically selecting misclassified negative samples during training. Training samples are heuristically sampled based on classification by the current status of the CNN. Weights are assigned to the training samples and informative samples are more likely to be included in the next CNN training iteration. We evaluated and compared our proposed method by training a CNN with (SeS) and without (NSeS) the selective sampling method. We focus on the detection of hemorrhages in color fundus images. A decreased training time from 170 epochs to 60 epochs with an increased performance-on par with two human experts-was achieved with areas under the receiver operating characteristics curve of 0.894 and 0.972 on two data sets. The SeS CNN statistically outperformed the NSeS CNN on an independent test set.
Patient-induced inhomogeneities in the magnetic field cause distortions and blurring during acquisitions with long readouts such as in susceptibility-weighted imaging (SWI). Most correction methods ...require collecting an additional ∆B 0 field map to remove these artifacts. The static ∆B0 field map can be approximated with an acceptable error directly from a single echo acquisition in SWI. The main component of the observed phase is linearly related to ∆B0 and the echo time TE , and the relative impact of non-∆B0 terms becomes insignificant with TE > 20ms at 3T for a well-tuned system. The main step is to combine and unfold the multi-channel phase maps wrapped many times, and several competing algorithms are compared for this purpose. Four in vivo brain data sets collected using the recently proposed 3D SPARKLING readouts are used to assess the proposed method. The estimated 3D field maps generated with a 0.6 mm isotropic spatial resolution provide overall similar off-resonance corrections compared to reference corrections based on an external ∆B0 acquisitions, and even improved for two out of four individuals. Although a small estimation error is expected, no aftermath was observed in the proposed corrections, while degradations were observed in the references. A static ∆B0 field map estimation method was proposed to take advantage of acquisitions with long echo times, and outperformed the reference technique based on an external field map. The difference can be attributed to an inherent robustness to mismatches between volumes and external ∆B0 maps, and diverse other sources investigated.
Most of ultrasound medical imaging systems currently on the market implement standard Delay and Sum (DAS) beamforming to form B-mode images. However, image resolution and contrast achievable with DAS ...are limited by the aperture size and by the operating frequency. For this reason, different beamformers have been presented in the literature that are mainly based on adaptive algorithms, which allow achieving higher performance at the cost of an increased computational complexity. In this paper, we propose the use of an alternative nonlinear beamforming algorithm for medical ultrasound imaging, which is called Delay Multiply and Sum (DMAS) and that was originally conceived for a RADAR microwave system for breast cancer detection. We modify the DMAS beamformer and test its performance on both simulated and experimentally collected linear-scan data, by comparing the Point Spread Functions, beampatterns, synthetic phantom and in vivo carotid artery images obtained with standard DAS and with the proposed algorithm. Results show that the DMAS beamformer outperforms DAS in both simulated and experimental trials and that the main improvement brought about by this new method is a significantly higher contrast resolution (i.e., narrower main lobe and lower side lobes), which turns out into an increased dynamic range and better quality of B-mode images.
The semantic image segmentation task consists of classifying each pixel of an image into an instance, where each instance corresponds to a class. This task is a part of the concept of scene ...understanding or better explaining the global context of an image. In the medical image analysis domain, image segmentation can be used for image-guided interventions, radiotherapy, or improved radiological diagnostics. In this review, we categorize the leading deep learning-based medical and non-medical image segmentation solutions into six main groups of deep architectural, data synthesis-based, loss function-based, sequenced models, weakly supervised, and multi-task methods and provide a comprehensive review of the contributions in each of these groups. Further, for each group, we analyze each variant of these groups and discuss the limitations of the current approaches and present potential future research directions for semantic image segmentation.
Radiomics is a relatively new word for the field of radiology, meaning the extraction of a high number of quantitative features from medical images. Artificial intelligence (AI) is broadly a set of ...advanced computational algorithms that basically learn the patterns in the data provided to make predictions on unseen data sets. Radiomics can be coupled with AI because of its better capability of handling a massive amount of data compared with the traditional statistical methods. Together, the primary purpose of these fields is to extract and analyze as much and meaningful hidden quantitative data as possible to be used in decision support. Nowadays, both radiomics and AI have been getting attention for their remarkable success in various radiological tasks, which has been met with anxiety by most of the radiologists due to the fear of replacement by intelligent machines. Considering ever-developing advances in computational power and availability of large data sets, the marriage of humans and machines in future clinical practice seems inevitable. Therefore, regardless of their feelings, the radiologists should be familiar with these concepts. Our goal in this paper was three-fold: first, to familiarize radiologists with the radiomics and AI; second, to encourage the radiologists to get involved in these ever-developing fields; and, third, to provide a set of recommendations for good practice in design and assessment of future works.
Introduction
Clinical utility and diagnostic sensitivity of new‐generation flat‐panel computed tomography perfusion imaging (FPCTP) performed immediately after mechanical thrombectomy (MT) is ...unknown. We aimed to assess whether FPCTP obtained directly after MT could provide additional potentially relevant information on tissue reperfusion status.
Methods
Qualitative, single‐center analysis of all consecutive acute stroke patients admitted between 06/2019 and 03/2021 who underwent MT and post‐interventional FPCTP (n=26). A core‐lab blinded to technical details and clinical data performed reperfusion grading on post‐interventional digital subtraction angiography (DSA) images and time‐sensitive FPCTP maps. All patients were classified according to agreement between DSA and FPCTP.
Results
In 10/26 patients FPCTP revealed new, potentially clinically relevant information. Core‐lab adjudicated dichotomized detection of hypoperfusion (present/absent) was concurring in 21/26 (81%) patients. Of these, reperfusion findings showed perfect agreement on location and size in 16 (62%) patients, while in 5 (19%) patients with incomplete reperfusion, FPCTP showed additional hypoperfused areas missed on DSA. Of the remaining five patients subject to disagreement regarding the presence or absence of hypoperfusion, three showed complete reperfusion on DSA but hypoperfusion was noted on FPCTP, whereas two showed incomplete reperfusion on DSA without detectable hypoperfusion on FPCTP. FPCTP findings could have avoided Thrombolysis in Cerebral Infarction (TICI) overestimation in all false‐positive operator‐rated TICI3 cases.
Conclusion
In both core‐lab and real‐world operator assessment, FPCTP may provide additional clinically relevant information in a considerable percentage of patients undergoing MT. Hence, FPCTP may constitute a new standard for evaluating reperfusion efficacy and informed decision making in the angiography suite.