John E. Ball

Digital Mammographic Computer Aided Diagnosis (CAD) using Adaptive Level Set Segmentation

We present a mammographic computer aided diagnosis (CAD) system, which uses an adaptive level set segmentation method (ALSSM), which segments suspicious masses in the polar domain and adaptively adjusts the border threshold at each angle to provide high-quality segmentation results. The primary contribution of this paper is the adaptive speed function for controlling level set segmentation. To assess the efficacy of the system, 60 relatively difficult cases (30 benign, 30 malignant) from the Digital Database of Screening Mammography (DDSM) are analyzed. The segmentation efficacy is analyzed qualitatively via visual inspection and quantitatively via the area under the receiver operating characteristics (ROC) curve (Az) and classification accuracies. For the ALSSM, the best results are 87% overall accuracy, Az=0.9687 with 28/30 malignant cases detected. The qualitative and quantitative results show that the ALSSM provides excellent segmentation and classification results and compares favorably to previous CAD systems in the literature which also used the DDSM database.

Digital Mammogram Spiculated Mass Detection and Spicule Segmentation using Level Sets

This letter presents an automated mammographic computer aided diagnosis (CAD) system to detect and segment spicules in digital mammograms, termed spiculation segmentation with level sets (SSLS). SSLS begins with a segmentation of the suspicious mass periphery, which is created using a previously developed adaptive level set segmentation algorithm (ALSSM) by the authors. The mammogram is then analyzed using features derived from the Dixon and Taylor line operator (DTLO), which is a method of linear structure enhancement. Features are extracted, optimized, and then the suspicious mass is classified as benign or malignant. To assess the system efficacy, 60 difficult mammographic images from the digital database of screening mammography (DDSM), containing 30 benign non-spiculated cases, 17 malignant spiculated cases, and 13 malignant non-spiculated cases, are analyzed. The initial spiculation detection method found 100% of the spiculated lesions with no false positive detections, and has area under the receiver operating characteristics (ROC) curve Az=1.0. The values using ALSSM (periphery segmentation only) are Az=0.9687 and 0.9708 for two investigated feature sets, and increases to Az=0.9862 using SSLS (spiculation segmentation). The best classification results are 93% overall accuracy (OA), with three false positives (FP) and one false negative (FN) using a 1-NN (nearest neighbor) or 2-NN classifier, and 92% OA with three FP and two FN using a maximum likelihood classifier.

Level Set Hyperspectral Image Classification Using Best Band Analysis

We present a supervised hyperspectral classification procedure consisting of an initial distance-based segmentation method that uses best band analysis (BBA), followed by a level set enhancement that forces localized region homogeneity. The proposed method is tested on two hyperspectral images of an urban and rural nature. The proposed method is compared to the maximum likelihood (ML) method using BBA. Quantitative results are compared using segmentation and classification accuracies. Results show that both the initial classification using BBA features and the level set enhancement produced high-quality ground cover maps and outperformed the ML method, as well as previous studies by the authors. For example, with the compact airborne spectrographic imager image, the ML method resulted in accuracies les95.5%, whereas the level set segmentation approach resulted in accuracies as high as 99.7%.

Level set segmentation of remotely sensed hyperspectral images

We present a semi-automated supervised hyperspectral image segmentation algorithm based on the level set methodology. In the proposed procedure, seed pixels are automatically selected by their similarity to the training signatures, and speed functions that control the level set propagation are created based on pixel similarity to the seed signature and class discriminator functions. Two sub images from a remotely sensed HYDICE hyperspectral image of the Washington D.C. Mall area in the U.S.A. are used to validate the algorithm. The results of the proposed algorithm are compared to the results using well-known supervised parallepiped or maximum-likelihood classification methods provided in the ERDAS Imagine software suite. The classes are grass, trees, buildings, water, paths and shadows. The results show the efficacy of the new algorithm. The contributions of the paper include: (1) successful application of the level set segmentation methodology to hyperspectral images, and (2) specification of speed functions suitable for controlling the level set propagation.

Hyperspectral Pixel Unmixing via Spectral Band Selection and DC-Insensitive Singular Value Decomposition

This letter proposes a linear two-class hyperspectral pixel-unmixing algorithm that uses a band selection method to determine the best bands for pixel unmixing, low-pass prefiltering to remove high-frequency content, and a new version of the well-known singular value decomposition (SVD) method, which is insensitive to dc offsets (DCI-SVD). The proposed method is compared to the best level discrete wavelet transform approach for dimensionality reduction and least squares estimation and quadratic programming for unmixing (DWT-LSE-QP). The contributions of this letter are given as follows: (1) the band selection and filter selection algorithm and (2) the DCI-SVD algorithm. The dc insensitivity of the DCI-SVD method is proven, and simulation results using data from an analytical spectral device spectroradiometer show the efficacy of the proposed method and its superiority to the DWT-LSE-QP-based approach in the harder unmixing cases.

Comprehensive survey of deep learning in remote sensing: theories, tools, and challenges for the community

Abstract. In recent years, deep learning (DL), a rebranding of neural networks (NNs), has risen to the top in numerous areas, namely computer vision (CV), speech recognition, and natural language processing. Whereas remote sensing (RS) possesses a number of unique challenges, primarily related to sensors and applications, inevitably RS draws from many of the same theories as CV, e.g., statistics, fusion, and machine learning, to name a few. This means that the RS community should not only be aware of advancements such as DL, but also be leading researchers in this area. Herein, we provide the most comprehensive survey of state-of-the-art RS DL research. We also review recent new developments in the DL field that can be used in DL for RS. Namely, we focus on theories, tools, and challenges for the RS community. Specifically, we focus on unsolved challenges and opportunities as they relate to (i) inadequate data sets, (ii) human-understandable solutions for modeling physical phenomena, (iii) big data, (iv) nontraditional heterogeneous data sources, (v) DL architectures and learning algorithms for spectral, spatial, and temporal data, (vi) transfer learning, (vii) an improved theoretical understanding of DL systems, (viii) high barriers to entry, and (ix) training and optimizing the DL.

Level set hyperspectral image segmentation using spectral information divergence-based best band selection

We present a supervised hyperspectral segmentation procedure, consisting of best band analysis (BBA), an initial distance-based segmentation, and level set segmentation enhancement by forcing localized vicinities to be more homogeneous. BBA uses the spectral information divergence (SID) to reduce each pixels high dimensional data to a scalar value, where the Bhattacharyya distance (BD) is maximized. The initial segmentation is based on feature vectors created from the SID metric. The level set segmentation then enhances areas that do not have spatially homogeneous ground cover. The proposed method is tested on a 72-band compact airborne spectrographic imager (CASI) image of a farm area in northern Mississippi, U.S.A. The proposed method is compared to a BBA-based maximum-likelihood (ML) method. Quantitative results are compared using segmentation and classification accuracies. Results show that both the initial classification using BBA features and the level set enhancement produced high-quality ground cover maps and outperformed the ML method, as well as previous studies by the authors. The ML method resulted in accuracies ges95.5%, whereas the level set segmentation approach resulted in accuracies as high as 99.7%.

Level Set Hyperspectral Segmentation: Near-Optimal Speed Functions using Best Band Analysis and Scaled Spectral Angle Mapper

This paper presents a semi-automated supervised level set hyperspectral image segmentation algorithm. The proposed method uses near-optimal speed functions (which control the level set segmentation) that are composed of a spectral similarity term and a stopping term. The spectral similarity term is used to compare pixels to class training signatures and is based on an optimized best bands analysis (BBA) procedure developed previously by the authors (2). The stopping term is created from a new BBA algorithm, which uses a modified version of the spectral angle mapper (SAM) called the scaled SAM (SSAM). The algorithm is validated with a HYDICE hyperspectral image of the Washington, D.C. Mall. The results of the proposed method are compared to previous results by the authors and show the efficacy of the new algorithm. The contributions of the paper include a nearly-optimal set of speed functions for hyperspectral level set analysis and an automated BBA algorithm based on the SSAM metric for creating the level set stopping term.

Hyperspectral pixel unmixing using singular value decomposition

A case study in pixel unmixing is performed using the singular value decomposition method. Using a linear mixing model, mixed pixels are created from a subset of the hyperspectral data from an analytical spectral devices handheld spectroradiometer. These pixels are unmixed using the other portion of the hyperspectral data. Simulation results are presented to illustrate the applicability of the presented technique.

Towards automated segmentation and classification of masses in mammograms

This work presents a straightforward approach to detecting and segmenting mammographic mass cores. The method utilizes adaptive thresholding applied to a contrast-enhanced version of the gray-scale mammogram, where the threshold is a function of the localized gray-level mean and variance. To assess the methods efficacy, it is applied to a database of 62 mammograms, each containing a suspicious mass (39 benign and 23 malignant). Each test case consists of a gray-scale image and a binary image containing a radiologist segmentation of the mass. After segmentation, a variety of features are extracted, including several based on the normalized radial length, rubber band straightening algorithm, gray-level statistics, and patient age. Next, step-wise linear discriminant analysis is utilized for feature reduction and optimization. The same procedure is applied to the manually segmented masses. Analysis of the optimized features resulted in an ROC curve area of Az = 0.8796 and Az = 0.8719 for the automated and manually segmented masses, respectively.