Pediatric tympanostomy tube assessment via deep learning.

PURPOSE: Tympanostomy tube (TT) placement is the most frequently performed ambulatory surgery in children under 15. After the procedure it is recommended that patients follow up regularly for "tube checks" until TT extrusion. Such visits incur direct and indirect costs to families in the form of days off from work, copays, and travel expenses. This pilot study aims to compare the efficacy of tympanic membrane (TM) evaluation by an artificial intelligence algorithm with that of clinical staff for determining presence or absence of a tympanostomy tube within the TM. METHODS: Using a digital otoscope, we performed a prospective study in children (ages 10 months-10 years) with a history of TTs who were being seen for follow up in a pediatric otolaryngology clinic. A smartphone otoscope was used by study personnel who were not physicians to take ear exam images, then through conventional otoscopic exam, ears were assessed by a clinician for tubes being in place or tubes having extruded from the TM. We trained and tested a deep learning (artificial intelligence) algorithm to assess the images and compared that with the clinician's assessment. RESULTS: A total of 123 images were obtained from 28 subjects. The algorithm classified images as TM with or without tube in place. Overall classification accuracy was 97.7 %. Recall and precision were 100 % and 96 %, respectively, for TM without a tube present, and 95 % and 100 %, respectively, for TM with a tube in place. DISCUSSION: This is a promising deep learning algorithm for classifying ear tube presence in the TM utilizing images obtained in awake children using an over-the-counter otoscope available to the lay population. We are continuing enrollment, with the goal of building an algorithm to assess tube patency and extrusion.

Herschel/HIFI observations of the circumstellar ammonia lines in IRC+10216.

CONTEXT: A discrepancy exists between the abundance of ammonia (NH3) derived previously for the circumstellar envelope (CSE) of IRC+10216 from far-IR submillimeter rotational lines and that inferred from radio inversion or mid-infrared (MIR) absorption transitions. AIMS: To address the discrepancy described above, new high-resolution far-infrared (FIR) observations of both ortho- and para-NH3 transitions toward IRC+10216 were obtained with Herschel, with the goal of determining the ammonia abundance and constraining the distribution of NH3 in the envelope of IRC+10216. METHODS: We used the Heterodyne Instrument for the Far Infrared (HIFI) on board Herschel to observe all rotational transitions up to the J = 3 level (three ortho- and six para-NH3 lines). We conducted non-LTE multilevel radiative transfer modelling, including the effects of near-infrared (NIR) radiative pumping through vibrational transitions. The computed emission line profiles are compared with the new HIFI data, the radio inversion transitions, and the MIR absorption lines in the ν2 band taken from the literature. RESULTS: We found that NIR pumping is of key importance for understanding the excitation of rotational levels of NH3. The derived NH3 abundances relative to molecular hydrogen were (2.8 ± 0.5) × 10-8 for ortho-NH3 and [Formula: see text] for para-NH3, consistent with an ortho/para ratio of 1. These values are in a rough agreement with abundances derived from the inversion transitions, as well as with the total abundance of NH3 inferred from the MIR absorption lines. To explain the observed rotational transitions, ammonia must be formed near to the central star at a radius close to the end of the wind acceleration region, but no larger than about 20 stellar radii (1σ confidence level).