Universal adversarial attacks on deep neural networks for medical image classification.

BACKGROUND: Deep neural networks (DNNs) are widely investigated in medical image classification to achieve automated support for clinical diagnosis. It is necessary to evaluate the robustness of medical DNN tasks against adversarial attacks, as high-stake decision-making will be made based on the diagnosis. Several previous studies have considered simple adversarial attacks. However, the vulnerability of DNNs to more realistic and higher risk attacks, such as universal adversarial perturbation (UAP), which is a single perturbation that can induce DNN failure in most classification tasks has not been evaluated yet. METHODS: We focus on three representative DNN-based medical image classification tasks (i.e., skin cancer, referable diabetic retinopathy, and pneumonia classifications) and investigate their vulnerability to the seven model architectures of UAPs. RESULTS: We demonstrate that DNNs are vulnerable to both nontargeted UAPs, which cause a task failure resulting in an input being assigned an incorrect class, and to targeted UAPs, which cause the DNN to classify an input into a specific class. The almost imperceptible UAPs achieved > 80% success rates for nontargeted and targeted attacks. The vulnerability to UAPs depended very little on the model architecture. Moreover, we discovered that adversarial retraining, which is known to be an effective method for adversarial defenses, increased DNNs' robustness against UAPs in only very few cases. CONCLUSION: Unlike previous assumptions, the results indicate that DNN-based clinical diagnosis is easier to deceive because of adversarial attacks. Adversaries can cause failed diagnoses at lower costs (e.g., without consideration of data distribution); moreover, they can affect the diagnosis. The effects of adversarial defenses may not be limited. Our findings emphasize that more careful consideration is required in developing DNNs for medical imaging and their practical applications.

Spectral domain optical coherence tomography classification of diabetic macular edema: a new proposal to clinical practice.

PURPOSE: To classify the types of diabetic macular edema (DME) and evaluate its morphological features on spectral domain optical coherence tomography (SD-OCT) and determine correlations between visual acuity and OCT findings. METHODS: We assessed 406 eyes of 309 patients with a diagnosis of DME retrospectively. Three types based on SD-OCT were identified: diffuse macular edema, cystoid macular edema, and cystoid degeneration. Morphological features such as serous macular detachment (SMD), vitreomacular interface abnormalities (VMAI), hard exudates, photoreceptor status, and correlations between visual acuity and those morphological features were also evaluated by SD-OCT. RESULTS: The most common type of DME was cystoid edema (68.5%). No statistically significant difference was found between groups in sex (P = 0.40), type of diabetes (P = 0.50), or diabetic retinopathy (P = 0.78). However, the duration of symptoms and BCVA was significantly lower in the group with cystoid degeneration compared with the group with cystoid edema (P < 0.001) and the group with diffuse macular edema (P < 0.001). In the group with cystoid degeneration compared with the groups with cystoid and diffuse edema, the central fovea and central subfield were significantly thicker (both (P < 0.001), the subfoveal choroid was significantly thinner (P = 0.049), rate of serous macular detachment was significantly lower (P < 0.001), and the rate of outer retinal damage was significantly higher (P < 0.001). CONCLUSIONS: Cystoid macular degeneration, which is consistent with poor functional and morphological outcomes, should be differentiated from cystoid macular edema. Serous macular detachment, which is mostly seen in eyes with early stages of DME, should be evaluated as an accompanying morphological finding rather than a type of DME.

DISEASE CLASSIFICATION OF MACULAR OPTICAL COHERENCE TOMOGRAPHY SCANS USING DEEP LEARNING SOFTWARE: Validation on Independent, Multicenter Data.

PURPOSE: To evaluate Pegasus optical coherence tomography (OCT), a clinical decision support software for the identification of features of retinal disease from macula OCT scans, across heterogenous populations involving varying patient demographics, device manufacturers, acquisition sites, and operators. METHODS: Five thousand five hundred and eighty-eight normal and anomalous macular OCT volumes (162,721 B-scans), acquired at independent centers in five countries, were processed using the software. Results were evaluated against ground truth provided by the data set owners. RESULTS: Pegasus-OCT performed with areas under the curve of the receiver operating characteristic of at least 98% for all data sets in the detection of general macular anomalies. For scans of sufficient quality, the areas under the curve of the receiver operating characteristic for general age-related macular degeneration and diabetic macular edema detection were found to be at least 99% and 98%, respectively. CONCLUSION: The ability of a clinical decision support system to cater for different populations is key to its adoption. Pegasus-OCT was shown to be able to detect age-related macular degeneration, diabetic macular edema, and general anomalies in OCT volumes acquired across multiple independent sites with high performance. Its use thus offers substantial promise, with the potential to alleviate the burden of growing demand in eye care services caused by retinal disease.

The design and validation of an optical coherence tomography-based classification system for focal vitreomacular traction.

PURPOSE: To develop and validate a classification system for focal vitreomacular traction (VMT) with and without macular hole based on spectral domain optical coherence tomography (SD-OCT), intended to aid in decision-making and prognostication. METHODS: A panel of retinal specialists convened to develop this system. A literature review followed by discussion on a wide range of cases formed the basis for the proposed classification. Key features on OCT were identified and analysed for their utility in clinical practice. A final classification was devised based on two sequential, independent validation exercises to improve interobserver variability. RESULTS: This classification tool pertains to idiopathic focal VMT assessed by a horizontal line scan using SD-OCT. The system uses width (W), interface features (I), foveal shape (S), retinal pigment epithelial changes (P), elevation of vitreous attachment (E), and inner and outer retinal changes (R) to give the acronym WISPERR. Each category is scored hierarchically. Results from the second independent validation exercise indicated a high level of agreement between graders: intraclass correlation ranged from 0.84 to 0.99 for continuous variables and Fleiss' kappa values ranged from 0.76 to 0.95 for categorical variables. CONCLUSIONS: We present an OCT-based classification system for focal VMT that allows anatomical detail to be scrutinised and scored qualitatively and quantitatively using a simple, pragmatic algorithm, which may be of value in clinical practice as well as in future research studies.

Classification of image artefacts in optical coherence tomography angiography of the choroid in macular diseases.

BACKGROUND: To evaluate and classify image artefacts in optical coherence tomography (OCT) angiography (OCTA) of the choroid in a group of patients with macular diseases. DESIGN: Retrospective observational study. PARTICIPANTS: Five patients with age-related macular degeneration, three with central serous retinopathy, one with polypoidal choroidal vasculopathy and one with multiple evanescent white dot syndrome. METHODS: OCTA and OCT reflectivity (OCTR) maps were reviewed along with their fluorescein angiography and indocyanine green angiography. Sixty OCTA images (20 outer retina, 20 Sattler and 20 Haller layers) were graded for image artefacts by two examiners independently. MAIN OUTCOME MEASURES: OCTA artefacts and their correlation with OCTR maps, angiography and OCT B-scans. RESULTS: Artefacts (frequency) were classified into (i) motion (70-100%), (ii) fringe washout (100%), (iii) decorrelation projection (0-20%), (iv) masking and unmasking (50-65%) and (v) stromal decorrelation signal (100%). Motion artefact in OCTA is characterized by horizontal dark lines or bands not apparent on OCTR map. Fringe washout creates signal void within choroidal vessels because of fast blood flow. Decorrelation projection from retinal vasculature and choroidal new vessels above the Bruch's membrane are seen within the choroidal OCTA image. Masking and unmasking artefacts occur in regions of pigment epithelial detachment and atrophy. Decorrelation signals can also be seen in the choroidal stroma. CONCLUSIONS: Our classification system of artefact in choroidal OCTA establishes a common terminology for clinical interpretation. This is important in enhancing our understanding of the principles of OCTA acquisition, and it also serves as a bench mark for reading centres.

RETINAL ASTROCYTIC HAMARTOMA: Optical Coherence Tomography Classification and Correlation With Tuberous Sclerosis Complex.

PURPOSE: To propose a classification of retinal astrocytic hamartoma based on spectral domain optical coherence tomography and correlate each class with systemic manifestations of tuberous sclerosis complex. METHODS: Retrospective chart review conducted at four international referral medical retina centers. There were 43 consecutive patients with an established diagnosis of tuberous sclerosis complex based on presence of at least 2 major or 1 major and 2 minor features of the diagnostic criteria. Clinical and spectral domain optical coherence tomography features regarding retinal astrocytic hamartoma were documented. RESULTS: The mean patient age at presentation was 16.2 years. The retinal astrocytic hamartoma was classified as Type I (n = 41), Type II (n = 25), Type III (n = 20), or Type IV (n = 12). Patients with Type II showed greater number of cutaneous fibrous plaques (odds ratio = 64.8; 92% confidence interval: 64.2-65; P < 0.001); those with Type III displayed higher incidence of subependymal giant-cell astrocytomas (odds ratio = 43.2; 95% confidence interval: 43.0-43.3; P < 0.001); and those with Type IV showed higher incidence of pulmonary lymphangiomyomatosis (odds ratio = 126; 95% confidence interval: 122-128; P < 0.001). CONCLUSION: Retinal astrocytic hamartoma can be classified into four morphologic groups, based on spectral domain optical coherence tomography. There are important systemic tuberous sclerosis complex correlations with each class.

SAVE: a grading protocol for clinically significant diabetic macular oedema based on optical coherence tomography and fluorescein angiography.

AIM: To analyse a new grading protocol for clinically significant diabetic macular oedema (CSME) based on spectral domain optical coherence tomography (SD-OCT) and fluorescein angiography (FA). METHODS: 56 eyes of 40 patients with CSME were examined by Cirrus OCT (Carl Zeiss Meditec), Spectralis HRA and OCT (Heidelberg Engineering) on the same day. Three graders analysed images based on a newly developed grading protocol integrating all relevant information from OCT and FA. The protocol defined four categories: (1) subretinal fluid (category 'S'); (2) the planimetrically measured oedematous area (category 'A'); (3) vitreo-retinal interface abnormalities (category 'V'); and (4) CSME aetiology (category 'E') defining the leakage source. RESULTS: The new grading protocol allowed for a detailed characterisation of each individual type of CSME. It defines four aetiological types of CSME and analyses four further categories important in diagnosis and during follow-up in clinical and study settings. Atrophic, a new type of CSME, was described and characteristic combinations of triggers of CSME were revealed. Inter-grader agreement, analysed using Fleiss' κ values for Cirrus OCT and Spectralis OCT, respectively, was good for 'S' (0.9; 0.82), 'A' (1.0; 1.0) and 'E' (range 0.63-0.8; 0.57-0.77), and lower for 'V' (0.25; 0.42). CONCLUSIONS: The 'SAVE' grading protocol of CSME integrates information from two imaging techniques, OCT and FA. Its clinical approach allows examiners to define and further categorise clinical characteristics to find tailored therapeutic strategies.

Determination of characteristics of degenerative joint disease using optical coherence tomography and polarization sensitive optical coherence tomography.

BACKGROUND AND OBJECTIVES: Previous studies have demonstrated that optical coherence tomography (OCT) could be used to delineate alterations in the microstructure of cartilage, and have suggested that changes in the polarization state of light as detected by OCT could provide information on the birefringence properties of articular cartilage as influenced by disease. In this study we have used both OCT and polarization sensitive optical coherence tomography (PS-OCT) technologies to evaluate normal and abnormal bovine articular cartilage according to established structural, organizational, and birefringent characteristics of degenerative joint disease (DJD) in order to determine if this technology can be used to differentiate various stages of DJD as a minimally invasive imaging tool. MATERIALS AND METHODS: Fresh bovine femoral-tibial joints were obtained from an abattoir, and 45 cartilage specimens were harvested from 8 tibial plateaus. Whole ex vivo specimens of normal and degenerative articular cartilage were imaged by both OCT and PS-OCT, then fixed and processed for histological evaluation. OCT/PS-OCT images and corresponding histology sections of each specimen were scored according to a modified Mankin structural grading scale and compared. RESULTS: OCT and PS-OCT imaging allowed structural evaluation of intact articular cartilage along a 6 mm surface length to a depth of 2 mm with a transverse resolution of 12 microm and an axial resolution of 10 microm. The OCT and PS-OCT images demonstrated characteristic alterations in the structure of articular cartilage with a high correlation to histological evaluation (kappa = 0.776). The OCT images were able to demonstrate early to advanced structural changes of articular cartilage while the optical phase retardation images obtained by PS-OCT imaging were able to discriminate areas where disorganization of the cartilage matrix was present, however, these characteristics are much different than those reported where OCT images alone were used to characterize tissue birefringence. No evidence of differences in OCT or PS-OCT images were detected between specimens of similar structural characteristics where proteoglycan was judged present or absent by safranin-O Fast Green staining. CONCLUSIONS: The combined use of OCT and PS-OCT technologies to obtain images from a single system is able to demonstrate and discriminate between characteristics of very early stages of surface irregularities not previously reported for OCT imaging, to deep clefts and collagen matrix disorganization for tissue at depths of up to 2 mm with good correlation to histology. PS-OCT and accumulated optical phase retardation images of articular cartilage as constructed from alterations in Stokes vector parameters appear to give a valuable but different assessment of alterations in tissue birefringence and organization than have been reported for OCT images obtained with the use of polarized or non-polarized light sources. This is the first time that alterations in the polarization state of light reflected from within the tissue have been demonstrated to be consistent with changes observed in the orientation and organization of the collagen matrix in advanced stages of DJD. The degree of phase transformation of light reflected from within the tissue as determined by PS-OCT imaging does not appear to be altered by the presence or absence of proteoglycan.