Machine Learning for Health: Algorithm Auditing & Quality Control.

Developers proposing new machine learning for health (ML4H) tools often pledge to match or even surpass the performance of existing tools, yet the reality is usually more complicated. Reliable deployment of ML4H to the real world is challenging as examples from diabetic retinopathy or Covid-19 screening show. We envision an integrated framework of algorithm auditing and quality control that provides a path towards the effective and reliable application of ML systems in healthcare. In this editorial, we give a summary of ongoing work towards that vision and announce a call for participation to the special issue  Machine Learning for Health: Algorithm Auditing & Quality Control in this journal to advance the practice of ML4H auditing.

Probability density function of the electric field in diffuse correlation spectroscopy of human bone in vivo.

Diffuse correlation spectroscopy (DCS) is the technique of choice for non-invasive assessments of human bone blood flow. However, DCS classical algorithms are based on the fundamental assumption that the electric field of the light reaching the DCS photodetector is a zero-mean complex Gaussian variable. The non-validity of this hypothesis might produce inaccurate blood flow estimations. It is shown that for the human tibia, the "Gaussian hypothesis" holds for interoptode distances ≥20 mm. This lower boundary seems to depend on the type of investigated tissue.

Absolute calibration of fiber-coupled single-photon detector.

We show a setup for characterising the efficiency of a single-photon-detector absolutely and with a precision better than 1%. Since the setup does not rely on calibrated devices and can be implemented with standard-optic components, it can be realised in any laboratory. Our approach is based on an Erbium-Doped-Fiber-Amplifier (EDFA) radiometer as a primary measurement standard for optical power, and on an ultra-stable source of spontaneous emission. As a proof of principle, we characterise the efficiency of an InGaAs/InP single-photon detector. We verified the correctness of the characterisation with independent measurements. In particular, the measurement of the optical power made with the EDFA radiometer has been compared to that of the Federal Institute of Metrology using a transfer power meter. Our approach is suitable for frequent characterisations of high-efficient single-photon detectors.

Time-domain algorithm for single-photon laser-Doppler flowmetry at large interoptode spacing in human bone.

A new laser-Doppler flowmeter at large interoptode spacing, based on single-photon counting (single-photon laser-Doppler flowmetry [SP-LDF]) and allowing assessment of blood flow deep in bone tissue, is proposed and implemented. To exploit the advantages of the new SP-LDF hardware, a dedicated simple and efficient time-domain algorithm has been developed. The new algorithm is based on the zero-order moment of the power density spectrum of the ad hoc prefiltered photoelectric current. The SP-LDF has been validated by Monte Carlo simulations, as well as by experimental measurements on a bone tissue phantom for optical flowmeters and on human.

MHz rate and efficient synchronous heralding of single photons at telecom wavelengths.

We report on the realization of a synchronous source of heralded single photons at telecom wavelengths with MHz heralding rates and high heralding efficiency. This source is based on the generation of photon pairs at 810 and 1550 nm via Spontaneous Parametric Down Conversion (SPDC) in a 1 cm periodically poled lithium niobate (PPLN) crystal pumped by a 532 nm pulsed laser. As high rates are fundamental for multi-photon experiments, we show that single telecom photons can be announced at 4.4 MHz rate with 45% heralding efficiency. When we focus only on the optimization of the coupling of the heralded photon, the heralding efficiency can be increased up to 80%. Furthermore, we experimentally observe that group velocity mismatch inside long crystals pumped in a pulsed mode affects the spectrum of the emitted photons and their fibre coupling efficiency. The length of the crystal in this source has been chosen as a trade off between high brightness and high coupling efficiency.

Advantages of gated silicon single-photon detectors.

We present a gated silicon single-photon detector based on a commercially available avalanche photodiode. Our detector achieves a photon-detection efficiency of 45±5% at 808 nm with 2·10(-6) dark count per nanosecond at 30 V of excess bias and -30°C. We compare gated and free-running detectors and show that this mode of operation has significant advantages in two representative experimental scenarios: detecting a single photon either hidden in faint continuous light or after a strong pulse. We also explore, at different temperatures and incident light intensities, the "charge persistence" effect, whereby a detector clicks some time after having been illuminated.

Statistical distortion of supervised learning predictions in optical microscopy induced by image compression.

The growth of data throughput in optical microscopy has triggered the extensive use of supervised learning (SL) models on compressed datasets for automated analysis. Investigating the effects of image compression on SL predictions is therefore pivotal to assess their reliability, especially for clinical use. We quantify the statistical distortions induced by compression through the comparison of predictions on compressed data to the raw predictive uncertainty, numerically estimated from the raw noise statistics measured via sensor calibration. Predictions on cell segmentation parameters are altered by up to 15% and more than 10 standard deviations after 16-to-8 bits pixel depth reduction and 10:1 JPEG compression. JPEG formats with higher compression ratios show significantly larger distortions. Interestingly, a recent metrologically accurate algorithm, offering up to 10:1 compression ratio, provides a prediction spread equivalent to that stemming from raw noise. The method described here allows to set a lower bound to the predictive uncertainty of a SL task and can be generalized to determine the statistical distortions originated from a variety of processing pipelines in AI-assisted fields.

Faithful entanglement swapping based on sum-frequency generation.

We show that an entanglement swapping operation performed with spontaneous parametric down-conversion can be made faithful without postselection using sum-frequency generation. This invites us to revisit the sum-frequency process and from a proof-of-principle experiment, we demonstrate that it provides a realistic solution for nonlinear optics at the single-photon level. This opens the way to attractive alternatives to six-photon protocols based on linear optics used, e.g., for the heralded creation of maximally entangled pairs or for device-independent quantum key distribution.

Room temperature photon number resolving detector for infared wavelengths.

In this paper we present a photon number resolving detector at infrared wavelengths, operating at room temperature and with a large dynamic range. It is based on the up-conversion of a signal at 1559 nm into visible wavelength and on its detection by a thermoelectrically cooled multi-pixel silicon avalanche photodiodode, also known as a Silicon Photon Multiplier. With the appropriate up-conversion this scheme can be implemented for arbitrary wavelengths above the visible spectral window. The preservation of the poissonian statistics when detecting coherent states is studied and the cross-talk effects on the detected signal can be easily estimated in order to calibrate the detector. This system is well suited for measuring very low intensities at infrared wavelengths and for analyzing multiphoton quantum states.

Quantum cloning for absolute radiometry.

In the quantum regime information can be copied with only a finite fidelity. This fidelity gradually increases to 1 as the system becomes classical. In this Letter we show how this fact can be used to directly measure the amount of radiated power. We demonstrate how these principles can be used to build a practical primary standard.

Bone tissue phantoms for optical flowmeters at large interoptode spacing generated by 3D-stereolithography.

A bone tissue phantom prototype allowing to test, in general, optical flowmeters at large interoptode spacings, such as laser-Doppler flowmetry or diffuse correlation spectroscopy, has been developed by 3D-stereolithography technique. It has been demonstrated that complex tissue vascular systems of any geometrical shape can be conceived. Absorption coefficient, reduced scattering coefficient and refractive index of the optical phantom have been measured to ensure that the optical parameters reasonably reproduce real human bone tissue in vivo. An experimental demonstration of a possible use of the optical phantom, utilizing a laser-Doppler flowmeter, is also presented.

Retinal and post-retinal contributions to the quantum efficiency of the human eye revealed by electrical neuroimaging.

The retina is one of the best known quantum detectors with rods able to reliably respond to single photons. However, estimates on the number of photons eliciting conscious perception, based on signal detection theory, are systematically above these values after discounting by retinal losses. One possibility is that there is a trade-off between the limited motor resources available to living systems and the excellent reliability of the visual photoreceptors. On this view, the limits to sensory thresholds are not set by the individual reliability of the receptors within each sensory modality (as often assumed) but rather by the limited central processing and motor resources available to process the constant inflow of sensory information. To investigate this issue, we reproduced the classical experiment from Hetch aimed to determine the sensory threshold in human vision. We combined a careful physical control of the stimulus parameters with high temporal/spatial resolution recordings of EEG signals and behavioral variables over a relatively large sample of subjects (12). Contrarily to the idea that the limits to visual sensitivity are fully set by the statistical fluctuations in photon absorption on retinal photoreceptors we observed that the state of ongoing neural oscillations before any photon impinges the retina helps to determine if the responses of photoreceptors have access to central conscious processing. Our results suggest that motivational and attentional off-retinal mechanisms play a major role in reducing the QE efficiency of the human visual system when compared to the efficiency of isolated retinal photoreceptors. Yet, this mechanism might subserve adaptive behavior by enhancing the overall multisensory efficiency of the whole system composed by diverse reliable sensory modalities.

32 bin near-infrared time-multiplexing detector with attojoule single-shot energy resolution.

We present two implementations of photon counting time-multiplexing detectors for near-infrared wavelengths, based on Peltier cooled InGaAs/InP avalanche photodiodes. A first implementation is motivated by practical considerations using only commercially available components. It features 16 bins, pulse repetition rates of up to 22 kHz, and a large range of applicable pulse widths of up to 100 ns. A second implementation is based on rapid gating detectors, permitting dead times below 10 ns. This allows one to realize a high dynamic-range 32 bin detector, able to process pulse repetition rates of up to 6 MHz for pulse widths of up to 200 ps. Analysis of the detector response at 16.5% detection efficiency reveals a single-shot energy resolution on the attojoule level.