A preprint comparing four reconstruction penalties on measured KIT4 electrical-impedance-tomography (EIT) data reports different trade-offs in the resulting images. L2-type reconstructions were generally smooth but diffuse, L1-type reconstructions produced cleaner backgrounds but sometimes distorted geometry, and smoothed total-variation (TV) and Huber-TV reconstructions showed more spatially coherent and broadly similar localization. In the final difficult test with two inclusions, however, none of the four penalties satisfactorily resolved both targets.
Inside the tank test
The study evaluated adaptive Nesterov momentum within a nonlinear complete-electrode-model (CEM) EIT reconstruction, using measured KIT4 data. The measurements came from a cylindrical saline tank 28 cm in diameter with 16 electrodes around its boundary. Each electrode was 7 cm high and 2.5 cm wide, and the data included 16 adjacent current-injection patterns at 2 mA. The evaluated phantoms covered conductive, insulating, mixed and geometrically challenging targets.
The essentially two-dimensional KIT4 geometry was discretized with finite elements on a mesh containing 7,132 triangular elements and 3,695 vertices. Each reconstruction iteration applied an adjoint-gradient correction, Nesterov momentum extrapolation and a dual-to-primal conductivity reconstruction. The data-misfit gradient was computed with CEM adjoint equations and stabilized using Sobolev smoothing.
Before reconstruction, background conductivity and contact impedance were calibrated from no-object measurements. The estimated contact impedance was then fixed for the reconstruction. The reported values were σbk = 1.8723 × 10−3 S and z = 2.5 × 10−4 Ω·cm.
The strong-convexity weight β was chosen empirically because no theoretical rule was available. L1 was more sensitive to β, particularly in whether the reconstructed support fragmented or shrank, while smoothed TV and Huber-TV were comparatively stable. The main experiments used βL1 = 5, βTVε = 2 and βTVγ = 2.
Different penalties, different images
In four single-inclusion configurations, L2 located all four cases but produced broad, diffuse variations. L1 gave the most homogeneous background, although target geometry could be distorted. Both TV variants localized the inclusions more accurately and were judged the best balance among the alternatives.
With two inclusions, the TV variants localized targets most accurately and preserved separation in the first three configurations. None of the four penalties satisfactorily resolved the final difficult configuration, which contained two targets.
In all-insulating multiple-target cases, all four methods identified the principal low-conductivity regions. When conductive and insulating targets appeared together, their opposite-sign contrasts remained distinguishable. The TV variants gave the most consistent separation, although their boundaries were smoother than the references and artifacts remained.
Triangular, pumpkin and hollow-cylinder phantoms were all localized near their correct positions by each method. L2 was more diffuse, while L1 and the TV variants were more compact or localized.
Where the comparison ran out of resolution
Large or boundary-adjacent inclusions were clearer. By contrast, small interior inclusions and closely spaced same-sign inclusions produced weak or overlapping responses.
The comparisons were visual and descriptive; no numerical accuracy score or resolution threshold was reported.
The document is marked arXiv:2608.25837v1 [math.NA] and dated 26 August 2026. Its evidence comes from measured saline-tank phantom data, making it a methods comparison for the tested KIT4 setup rather than evidence of clinical performance or a universal ranking of penalties.
Paper data and sources
Original title: Adaptive Nesterov Momentum Method for Electrical Impedance Tomography with the Complete Electrode Model
Authors: Kai Zhu, Jijun Liu, Min Zhong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text