Siegfried Wahl

Leveraging uncertainty information from deep neural networks for disease detection

Deep learning (DL) has revolutionized the field of computer vision and image processing. In medical imaging, algorithmic solutions based on DL have been shown to achieve high performance on tasks that previously required medical experts. However, DL-based solutions for disease detection have been proposed without methods to quantify and control their uncertainty in a decision. In contrast, a physician knows whether she is uncertain about a case and will consult more experienced colleagues if needed. Here we evaluate drop-out based Bayesian uncertainty measures for DL in diagnosing diabetic retinopathy (DR) from fundus images and show that it captures uncertainty better than straightforward alternatives. Furthermore, we show that uncertainty informed decision referral can improve diagnostic performance. Experiments across different networks, tasks and datasets show robust generalization. Depending on network capacity and task/dataset difficulty, we surpass 85% sensitivity and 80% specificity as recommended by the NHS when referring 0−20% of the most uncertain decisions for further inspection. We analyse causes of uncertainty by relating intuitions from 2D visualizations to the high-dimensional image space. While uncertainty is sensitive to clinically relevant cases, sensitivity to unfamiliar data samples is task dependent, but can be rendered more robust.

The Influence of Induced Astigmatism on the Depth of Focus.

Purpose To evaluate whether an induced astigmatism influences the subjective depth of focus. Methods Fifty-one participants aged 18 to 35 years and with a mean spherical equivalent refractive error of −0.51 ± 2.35 DS participated in the study. The accommodation was blocked with three drops of 1% cyclopentolate. Refractive errors were corrected after subjective refraction with a 4-mm artificial pupil. To evaluate the depth of focus (DoF), defocus curves with a spherical range of ±1.5 DS were assessed. The DoF was calculated as the horizontal distance at a threshold level of +0.1 logMAR from the maximum visual acuity (VA). Defocus curves were estimated binocularly during distance (500 cm) and a near vision (40 cm) for two induced axis (ATR in 0° and WTR in 90°) and for a fixed amount of astigmatic defocus of −0.5 DC. Results The mean natural DoF was 0.885 ± 0.316 D for far vision and 0.940 ± 0.400 D for near vision. With induced astigmatism, the DoF for far vision was significantly increased to 1.095 ± 0.421 D (p = 0.006, ANOVA) for the WTR astigmatism but not for the ATR astigmatism (1.030 ± 0.395 D; p = 0.164, ANOVA). The induced WTR astigmatism enhanced the DoF for near vision significantly to 1.144 ± 0.338 D (p = 0.04, ANOVA), and DoF with induced ATR astigmatism (0.953 ± 0.318 D) was not significantly different (p = 1.00, ANOVA). ATR-astigmatism reduced VA by +0.08 ± 0.08 logMAR (p < 0.01, t-test). Conclusions With an induced WTR astigmatism of −0.5 DC, the DoF can be enhanced in the near viewing distance with a marginal loss in binocular VA. The approach of using induced WTR astigmatism can lead to novel optical treatments for presbyopia.

Distorted Low-Level Visual Features Affect Saliency-Based Visual Attention

Image distortions can attract attention away from the natural scene saliency (Redi et al., 2011). Performance of viewers in visual search tasks and their fixation patterns are also affected by different types and amounts of distortions (Vu et al., 2008). In this paper, we have discussed the opinion that distortions could largely affect the performance of predictive models of visual attention, and simulated the effects of distorted low-level visual features on the saliency-based bottom-up visual attention. Saliency is a fast and pre-attentive mechanism for orienting visual attention to intrinsically important objects which pop-out more easily in a cluttered scene. Distortion of the low-level features that contribute to saliency may impair the readiness of the visual system in detection of salient objects, which may have major implications for critical situations like driving or locomotion. These distortions in natural life can be introduced by eye diseases such as cataract, or spectacles which may alter color perception (de Fez et al., 2002) or cause undesired optical effects like blurring, non-uniform magnification, and image displacement (Barbero and Portilla, 2016). The extent to which each of these distorted saliency features may affect the attentional performance is addressed in this paper by employing a biologically-inspired predictive model of visual attention. We briefly summarize the current standing of computational work on visual attention models in the following section and suggest a simple and influential model of saliency to examine the above hypothesis. Furthermore, we demonstrate in an example the hindered performance of the predictive saliency model on distorted images.

Lack of oblique astigmatism in the chicken eye

Primate eyes display considerable oblique off-axis astigmatism which could provide information on the sign of defocus that is needed for emmetropization. The pattern of peripheral astigmatism is not known in the chicken eye, a common model of myopia. Peripheral astigmatism was mapped out over the horizontal visual field in three chickens, 43 days old, and in three near emmetropic human subjects, average age 34.7years, using infrared photoretinoscopy. There were no differences in astigmatism between humans and chickens in the central visual field (chicks -0.35D, humans -0.65D, n.s.) but large differences in the periphery (i.e. astigmatism at 40° in the temporal visual field: humans -4.21D, chicks -0.63D, p<0.001, unpaired t-test). The lack of peripheral astigmatism in chicks was not due to differences in corneal shape. Perhaps related to their superior peripheral optics, we found that chickens had excellent visual performance also in the far periphery. Using an automated optokinetic nystagmus paradigm, no difference was observed in spatial visual performance with vision restricted to either the central 67° of the visual field or to the periphery beyond 67°. Accommodation was elicited by stimuli presented far out in the visual field. Transscleral images of single infrared LEDs showed no sign of peripheral astigmatism. The chick may be the first terrestrial vertebrate described to lack oblique astigmatism. Since corneal shape cannot account for the difference in astigmatism in humans and chicks, it must trace back to the design of the crystalline lens. The lack of peripheral astigmatism in chicks also excludes a role in emmetropization.

TuebingenCSTest – a useful method to assess the contrast sensitivity function

Since contrast sensitivity (CS) relies on the accuracy of stimulus presentation, the reliability of the psychophysical procedure and observers attention, the measurement of the CS-function is critical and therefore, a useful threshold contrast measurement was developed. The Tuebingen Contrast Sensitivity Test (TueCST) includes an adaptive staircase procedure and a 16-bit gray-level resolution. In order to validate the CS measurements with the TueCST, measurements were compared with existing tests by inter-test repeatability, test-retest reliability and time. The novel design enables an accurate presentation of the spatial frequency and higher precision, inter-test repeatability and test-retest reliability compared to other existing tests.

Can positions in the visual field with high attentional capabilities be good candidates for a new preferred retinal locus

&NA; The sustained component of visual attention lowers the perceptual threshold of stimuli located at the attended region. Attentional performance is not equal for all eccentric positions, leading to variations in perception. The location of the preferred retinal locus (PRL) for fixation might be influenced by these attentional variations. This study investigated the relation between the placement of sustained attention and the location of a developed PRL using simulations of central scotoma. Thirteen normally sighted subjects participated in the study. Monocular sustained attention was measured in discrete eccentric locations of the visual field using the dominant eye. Subsequently, a six degrees macular scotoma was simulated and PRL training was performed during eight ten‐minutes blocks of trials. After training, every subject developed a PRL. Subjects with high attentional capabilities in the lower hemifield generally developed PRLs in the lower hemifield (n = 10), subjects with high attentional capabilities in the upper hemifield developed PRLs in the upper hemifield (n = 2) and one subject with similar attentional capabilities in the upper and lower hemifield developed the PRL on the upper hemifield. Analyzed individually, the results showed that 70% of the subjects had a PRL location in the hemifield where high attentional performance was achieved. These results suggest that attentional capabilities can be used as a predictor for the development of the PRL and are of significance for low vision rehabilitation and for the development of new PRL training procedures, with the option for a preventive attentional training in early macular disease to develop a favorable PRL.

A preferred retinal location of fixation can be induced when systematic stimulus relocations are applied

Patients with central vision loss obtain visual information by fixating on an object eccentrically with a preferred retinal locus of fixation (PRL). Patients do not always choose the most efficient PRL position, and as a consequence, visual performance is not always fully exploited. This study investigates whether PRLs can be induced by applying systematic stimulus relocations. The PRL was trained using a central scotoma simulation in 15 healthy subjects. They performed different visual tasks during four sessions, after which their reading performance was evaluated. In five subjects the stimulus was relocated to the left hemifield whenever a saccade would place the stimulus on the opposite hemifield. In five different subjects the relocation was inversed: The stimulus was located in the right hemifield. The relocation was 7.5° of visual angle and it was applied horizontally. Five additional subjects naturally chose the PRL location. They were used as the control group to evaluate the development of a PRL. After training, subjects performed visual search tasks on static stimuli. Evaluation after training showed that systematic stimulus relocations can be used to influence the development of the PRL. These results might be significant for the development of training strategies for the visually impaired.

Adaptation to Skew Distortions of Natural Scenes and Retinal Specificity of Its Aftereffects

Image skew is one of the prominent distortions that exist in optical elements, such as in spectacle lenses. The present study evaluates adaptation to image skew in dynamic natural images. Moreover, the cortical levels involved in skew coding were probed using retinal specificity of skew adaptation aftereffects. Left and right skewed natural image sequences were shown to observers as adapting stimuli. The point of subjective equality (PSE), i.e., the skew amplitude in simple geometrical patterns that is perceived to be unskewed, was used to quantify the aftereffect of each adapting skew direction. The PSE, in a two-alternative forced choice paradigm, shifted toward the adapting skew direction. Moreover, significant adaptation aftereffects were obtained not only at adapted, but also at non-adapted retinal locations during fixation. Skew adaptation information was transferred partially to non-adapted retinal locations. Thus, adaptation to skewed natural scenes induces coordinated plasticity in lower and higher cortical areas of the visual pathway.

Non-intrusive practitioner pupil detection for unmodified microscope oculars

Modern microsurgery is a long and complex task requiring the surgeon to handle multiple microscope controls while performing the surgery. Eye tracking provides an additional means of interaction for the surgeon that could be used to alleviate this situation, diminishing surgeon fatigue and surgery time, thus decreasing risks of infection and human error. In this paper, we introduce a novel algorithm for pupil detection tailored for eye images acquired through an unmodified microscope ocular. The proposed approach, the Hough transform, and six state-of-the-art pupil detection algorithms were evaluated on over 4000 hand-labeled images acquired from a digital operating microscope with a non-intrusive monitoring system for the surgeon eyes integrated. Our results show that the proposed method reaches detection rates up to 71% for an error of ≈3% w.r.t the input image diagonal; none of the state-of-the-art pupil detection algorithms performed satisfactorily. The algorithm and hand-labeled data set can be downloaded at:: www.ti.uni-tuebingen.de/perception.

Transsaccadic transfer of distortion adaptation in a natural environment

Spatially varying distortions in optical elements-for instance prisms and progressive power lenses-modulate the visual world disparately in different visual areas. Saccadic eye movements in such a complexly distorted environment thereby continuously alter the retinal location of the distortions. Yet the visual system achieves perceptual constancy by compensating for distortions irrespective of their retinal relocations at different fixations. Here, we assessed whether the visual system retains its plasticity to distortions across saccades to attain stability. Specifically, we tapped into reference frames of geometric skew-adaptation aftereffects to evaluate the transfer of retinotopic and spatiotopic distortion information across saccades. Adaptation to skew distortion of natural-image content was tested at retinotopic and spatiotopic locations after a saccade was executed between adaptation and test phases. The skew-adaptation information was partially transferred to a new fixation after a saccade. Significant adaptation aftereffects were obtained at both retinotopic and spatiotopic locations. Conceivably, spatiotopic information was used to counterbalance the saccadic retinal shifts of the distortions. Therefore, distortion processing in a natural visual world does not start anew at each fixation; rather, retinotopic and spatiotopic skew information acquired at previous fixations are preserved to mediate stable perception during eye movements.