Holger R. Roth

Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning

Remarkable progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets and deep convolutional neural networks (CNNs). CNNs enable learning data-driven, highly representative, hierarchical image features from sufficient training data. However, obtaining datasets as comprehensively annotated as ImageNet in the medical imaging domain remains a challenge. There are currently three major techniques that successfully employ CNNs to medical image classification: training the CNN from scratch, using off-the-shelf pre-trained CNN features, and conducting unsupervised CNN pre-training with supervised fine-tuning. Another effective method is transfer learning, i.e., fine-tuning CNN models pre-trained from natural image dataset to medical image tasks. In this paper, we exploit three important, but previously understudied factors of employing deep convolutional neural networks to computer-aided detection problems. We first explore and evaluate different CNN architectures. The studied models contain 5 thousand to 160 million parameters, and vary in numbers of layers. We then evaluate the influence of dataset scale and spatial image context on performance. Finally, we examine when and why transfer learning from pre-trained ImageNet (via fine-tuning) can be useful. We study two specific computer-aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification. We achieve the state-of-the-art performance on the mediastinal LN detection, and report the first five-fold cross-validation classification results on predicting axial CT slices with ILD categories. Our extensive empirical evaluation, CNN model analysis and valuable insights can be extended to the design of high performance CAD systems for other medical imaging tasks.Remarkable progress has been made in image recognition, primarily due to the availability of large-scale annotated datasets and deep convolutional neural networks (CNNs). CNNs enable learning data-driven, highly representative, hierarchical image features from sufficient training data. However, obtaining datasets as comprehensively annotated as ImageNet in the medical imaging domain remains a challenge. There are currently three major techniques that successfully employ CNNs to medical image classification: training the CNN from scratch, using off-the-shelf pre-trained CNN features, and conducting unsupervised CNN pre-training with supervised fine-tuning. Another effective method is transfer learning, i.e., fine-tuning CNN models pre-trained from natural image dataset to medical image tasks. In this paper, we exploit three important, but previously understudied factors of employing deep convolutional neural networks to computer-aided detection problems. We first explore and evaluate different CNN architectures. The studied models contain 5 thousand to 160 million parameters, and vary in numbers of layers. We then evaluate the influence of dataset scale and spatial image context on performance. Finally, we examine when and why transfer learning from pre-trained ImageNet (via fine-tuning) can be useful. We study two specific computer-aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification. We achieve the state-of-the-art performance on the mediastinal LN detection, and report the first five-fold cross-validation classification results on predicting axial CT slices with ILD categories. Our extensive empirical evaluation, CNN model analysis and valuable insights can be extended to the design of high performance CAD systems for other medical imaging tasks.

A New 2.5D Representation for Lymph Node Detection Using Random Sets of Deep Convolutional Neural Network Observations

Automated Lymph Node (LN) detection is an important clinical diagnostic task but very challenging due to the low contrast of surrounding structures in Computed Tomography (CT) and to their varying sizes, poses, shapes and sparsely distributed locations. State-of-the-art studies show the performance range of 52.9% sensitivity at 3.1 false-positives per volume (FP/vol.), or 60.9% at 6.1 FP/vol. for mediastinal LN, by one-shot boosting on 3D HAAR features. In this paper, we first operate a preliminary candidate generation stage, towards -100% sensitivity at the cost of high FP levels (-40 per patient), to harvest volumes of interest (VOI). Our 2.5D approach consequently decomposes any 3D VOI by resampling 2D reformatted orthogonal views N times, via scale, random translations, and rotations with respect to the VOI centroid coordinates. These random views are then used to train a deep Convolutional Neural Network (CNN) classifier. In testing, the CNN is employed to assign LN probabilities for all N random views that can be simply averaged (as a set) to compute the final classification probability per VOI. We validate the approach on two datasets: 90 CT volumes with 388 mediastinal LNs and 86 patients with 595 abdominal LNs. We achieve sensitivities of 70%/83% at 3 FP/vol. and 84%/90% at 6 FP/vol. in mediastinum and abdomen respectively, which drastically improves over the previous state-of-the-art work.

Improving Computer-Aided Detection Using Convolutional Neural Networks and Random View Aggregation

Automated computer-aided detection (CADe) has been an important tool in clinical practice and research. State-of-the-art methods often show high sensitivities at the cost of high false-positives (FP) per patient rates. We design a two-tiered coarse-to-fine cascade framework that first operates a candidate generation system at sensitivities ~ 100% of but at high FP levels. By leveraging existing CADe systems, coordinates of regions or volumes of interest (ROI or VOI) are generated and function as input for a second tier, which is our focus in this study. In this second stage, we generate 2D (two-dimensional) or 2.5D views via sampling through scale transformations, random translations and rotations. These random views are used to train deep convolutional neural network (ConvNet) classifiers. In testing, the ConvNets assign class (e.g., lesion, pathology) probabilities for a new set of random views that are then averaged to compute a final per-candidate classification probability. This second tier behaves as a highly selective process to reject difficult false positives while preserving high sensitivities. The methods are evaluated on three data sets: 59 patients for sclerotic metastasis detection, 176 patients for lymph node detection, and 1,186 patients for colonic polyp detection. Experimental results show the ability of ConvNets to generalize well to different medical imaging CADe applications and scale elegantly to various data sets. Our proposed methods improve performance markedly in all cases. Sensitivities improved from 57% to 70%, 43% to 77%, and 58% to 75% at 3 FPs per patient for sclerotic metastases, lymph nodes and colonic polyps, respectively.

Spatial Aggregation of Holistically-Nested Networks for Automated Pancreas Segmentation

Accurate automatic organ segmentation is an important yet challenging problem for medical image analysis. The pancreas is an abdominal organ with very high anatomical variability. This inhibits traditional segmentation methods from achieving high accuracies, especially compared to other organs such as the liver, heart or kidneys. In this paper, we present a holistic learning approach that integrates semantic mid-level cues of deeply-learned organ interior and boundary maps via robust spatial aggregation using random forest. Our method generates boundary preserving pixel-wise class labels for pancreas segmentation. Quantitative evaluation is performed on CT scans of 82 patients in 4-fold cross-validation. We achieve a (mean ± std. dev.) Dice Similarity Coefficient of 78.01 %±8.2 % in testing which significantly outperforms the previous state-of-the-art approach of 71.8 %±10.7 % under the same evaluation criterion.

Holistic classification of CT attenuation patterns for interstitial lung diseases via deep convolutional neural networks

Interstitial lung diseases (ILD) involve several abnormal imaging patterns observed in computed tomography (CT) images. Accurate classification of these patterns plays a significant role in precise clinical decision making of the extent and nature of the diseases. Therefore, it is important for developing automated pulmonary computer-aided detection systems. Conventionally, this task relies on experts’ manual identification of regions of interest (ROIs) as a prerequisite to diagnose potential diseases. This protocol is time consuming and inhibits fully automatic assessment. In this paper, we present a new method to classify ILD imaging patterns on CT images. The main difference is that the proposed algorithm uses the entire image as a holistic input. By circumventing the prerequisite of manual input ROIs, our problem set-up is significantly more difficult than previous work but can better address the clinical workflow. Qualitative and quantitative results using a publicly available ILD database demonstrate state-of-the-art classification accuracy under the patch-based classification and shows the potential of predicting the ILD type using holistic image.

Anatomy-specific classification of medical images using deep convolutional nets

Automated classification of human anatomy is an important prerequisite for many computer-aided diagnosis systems. The spatial complexity and variability of anatomy throughout the human body makes classification difficult. “Deep learning” methods such as convolutional networks (ConvNets) outperform other state-of-the-art methods in image classification tasks. In this work, we present a method for organ- or body-part-specific anatomical classification of medical images acquired using computed tomography (CT) with ConvNets. We train a ConvNet, using 4,298 separate axial 2D key-images to learn 5 anatomical classes. Key-images were mined from a hospital PACS archive, using a set of 1,675 patients. We show that a data augmentation approach can help to enrich the data set and improve classification performance. Using ConvNets and data augmentation, we achieve anatomy-specific classification error of 5.9 % and area-under-the-curve (AUC) values of an average of 0.998 in testing. We demonstrate that deep learning can be used to train very reliable and accurate classifiers that could initialize further computer-aided diagnosis.

Detection of Sclerotic Spine Metastases via Random Aggregation of Deep Convolutional Neural Network Classifications

Automated detection of sclerotic metastases (bone lesions) in Computed Tomography (CT) images has potential to be an important tool in clinical practice and research. State-of-the-art methods show performance of 79 % sensitivity or true-positive (TP) rate, at 10 false-positives (FP) per volume. We design a two-tiered coarse-to-fine cascade framework to first operate a highly sensitive candidate generation system at a maximum sensitivity of \(\sim \)92 % but with high FP level (\(\sim \)50 per patient). Regions of interest (ROI) for lesion candidates are generated in this step and function as input for the second tier. In the second tier we generate \(N\) 2D views, via scale, random translations, and rotations with respect to each ROI centroid coordinates. These random views are used to train a deep Convolutional Neural Network (CNN) classifier. In testing, the CNN is employed to assign individual probabilities for a new set of \(N\) random views that are averaged at each ROI to compute a final per-candidate classification probability. This second tier behaves as a highly selective process to reject difficult false positives while preserving high sensitivities. We validate the approach on CT images of 59 patients (49 with sclerotic metastases and 10 normal controls). The proposed method reduces the number of FP/vol. from 4 to 1.2, 7 to 3, and 12 to 9.5 when comparing a sensitivity rates of 60, 70, and 80 % respectively in testing. The Area-Under-the-Curve (AUC) is 0.834. The results show marked improvement upon previous work.

Registration of the endoluminal surfaces of the colon derived from prone and supine CT colonography

PURPOSE Computed tomographic (CT) colonography is a relatively new technique for detecting bowel cancer or potentially precancerous polyps. CT scanning is combined with three-dimensional (3D) image reconstruction to produce a virtual endoluminal representation similar to optical colonoscopy. Because retained fluid and stool can mimic pathology, CT data are acquired with the bowel cleansed and insufflated with gas and patient in both prone and supine positions. Radiologists then match visually endoluminal locations between the two acquisitions in order to determine whether apparent pathology is real or not. This process is hindered by the fact that the colon, essentially a long tube, can undergo considerable deformation between acquisitions. The authors present a novel approach to automatically establish spatial correspondence between prone and supine endoluminal colonic surfaces after surface parameterization, even in the case of local colon collapse. METHODS The complexity of the registration task was reduced from a 3D to a 2D problem by mapping the surfaces extracted from prone and supine CT colonography onto a cylindrical parameterization. A nonrigid cylindrical registration was then performed to align the full colonic surfaces. The curvature information from the original 3D surfaces was used to determine correspondence. The method can also be applied to cases with regions of local colonic collapse by ignoring the collapsed regions during the registration. RESULTS Using a development set, suitable parameters were found to constrain the cylindrical registration method. Then, the same registration parameters were applied to a different set of 13 validation cases, consisting of 8 fully distended cases and 5 cases exhibiting multiple colonic collapses. All polyps present were well aligned, with a mean (+/- std. dev.) registration error of 5.7 (+/- 3.4) mm. An additional set of 1175 reference points on haustral folds spread over the full endoluminal colon surfaces resulted in an error of 7.7 (+/- 7.4) mm. Here, 82% of folds were aligned correctly after registration with a further 15% misregistered by just onefold. CONCLUSIONS The proposed method reduces the 3D registration task to a cylindrical registration representing the endoluminal surface of the colon. Our algorithm uses surface curvature information as a similarity measure to drive registration to compensate for the large colorectal deformations that occur between prone and supine data acquisitions. The method has the potential to both enhance polyp detection and decrease the radiologists interpretation time.

A Bottom-Up Approach for Pancreas Segmentation Using Cascaded Superpixels and (Deep) Image Patch Labeling

Robust organ segmentation is a prerequisite for computer-aided diagnosis, quantitative imaging analysis, pathology detection, and surgical assistance. For organs with high anatomical variability (e.g., the pancreas), previous segmentation approaches report low accuracies, compared with well-studied organs, such as the liver or heart. We present an automated bottom-up approach for pancreas segmentation in abdominal computed tomography (CT) scans. The method generates a hierarchical cascade of information propagation by classifying image patches at different resolutions and cascading (segments) superpixels. The system contains four steps: 1) decomposition of CT slice images into a set of disjoint boundary-preserving superpixels; 2) computation of pancreas class probability maps via dense patch labeling; 3) superpixel classification by pooling both intensity and probability features to form empirical statistics in cascaded random forest frameworks; and 4) simple connectivity based post-processing. Dense image patch labeling is conducted using two methods: efficient random forest classification on image histogram, location and texture features; and more expensive (but more accurate) deep convolutional neural network classification, on larger image windows (i.e., with more spatial contexts). Over-segmented 2-D CT slices by the simple linear iterative clustering approach are adopted through model/parameter calibration and labeled at the superpixel level for positive (pancreas) or negative (non-pancreas or background) classes. The proposed method is evaluated on a data set of 80 manually segmented CT volumes, using six-fold cross-validation. Its performance equals or surpasses other state-of-the-art methods (evaluated by “leave-one-patient-out”), with a dice coefficient of 70.7% and Jaccard index of 57.9%. In addition, the computational efficiency has improved significantly, requiring a mere 6 ~ 8 min per testing case, versus ≥ 10 h for other methods. The segmentation framework using deep patch labeling confidences is also more numerically stable, as reflected in the smaller performance metric standard deviations. Finally, we implement a multi-atlas label fusion (MALF) approach for pancreas segmentation using the same data set. Under six-fold cross-validation, our bottom-up segmentation method significantly outperforms its MALF counterpart: 70.7±13.0% versus 52.51±20.84% in dice coefficients.

2D View Aggregation for Lymph Node Detection Using a Shallow Hierarchy of Linear Classifiers

Enlarged lymph nodes (LNs) can provide important information for cancer diagnosis, staging, and measuring treatment reactions, making automated detection a highly sought goal. In this paper, we propose a new algorithm representation of decomposing the LN detection problem into a set of 2D object detection subtasks on sampled CT slices, largely alleviating the curse of dimensionality issue. Our 2D detection can be effectively formulated as linear classification on a single image feature type of Histogram of Oriented Gradients (HOG), covering a moderate field-of-view of 45 by 45 voxels. We exploit both max-pooling and sparse linear fusion schemes to aggregate these 2D detection scores for the final 3D LN detection. In this manner, detection is more tractable and does not need to perform perfectly at instance level (as weak hypotheses) since our aggregation process will robustly harness collective information for LN detection. Two datasets (90 patients with 389 mediastinal LNs and 86 patients with 595 abdominal LNs) are used for validation. Cross-validation demonstrates 78.0% sensitivity at 6 false positives/volume (FP/vol.) (86.1% at 10 FP/vol.) and 73.1% sensitivity at 6 FP/vol. (87.2% at 10 FP/vol.), for the mediastinal and abdominal datasets respectively. Our results compare favorably to previous state-of-the-art methods.