Neural Noise

DTCNet: finger flexion decoding with three-dimensional ECoG data

3 min read

  • DTCNet is a decoder architecture for finger flexion decoding from three-dimensional ECoG data.1
  • Three-dimensional spatial ECoG decoding supports ECoG-based motor BCI for neuroprosthetics and neurorobotics.1 1

Gardner updates

  • DTCNet is a decoder architecture for finger flexion decoding from three-dimensional ECoG data, supporting ECoG-based motor BCI for neuroprosthetics and neurorobotics (Frontiers). 1

Weekly enrichment (2026-07-20)

  • DTCNet was published in Frontiers in Computational Neuroscience (Volume 19, 2025; DOI 10.3389/fncom.2025.1627819) and reframes ECoG finger decoding by converting 2D multi-channel ECoG into 3D spatio-temporal spectrograms (electrode channel × wavelet frequency × time) via a Morlet wavelet transform.2 3
  • On the public BCI Competition IV Dataset 4 (three human subjects, S1–S3), DTCNet became the first method to push the correlation between predicted and actual multi-finger flexion trajectories above 80%, reaching a peak correlation coefficient of 0.82 (thumb, subject S3).2 3
  • Averaged across the thumb, index, middle, ring, and little fingers and all three subjects, DTCNet achieved a mean Pearson correlation of 0.69, a 2.98% improvement over prior state-of-the-art methods on the same benchmark.2 3
  • The source ECoG was recorded with the BCI2000 system, band-pass filtered 0.15–200 Hz at a 1,000 Hz sampling rate, while finger bending angles were captured at 25 Hz with a data glove; both streams were resampled to a common 100 Hz for training.3
  • Task design cued each finger movement for 2 s (subjects repeated the cued finger movement three to five times) followed by a 2 s blank-screen rest; preprocessing added a 40–300 Hz band-pass and a 60 Hz notch filter to suppress physiological and power-line noise.3
  • The architecture is an encoder–decoder using 1D dilated convolutions (kernel sizes 7,7,5,5,5 with a sawtooth dilation pattern 1,2,3,1,2) for feature extraction and transposed convolutions with skip connections for temporal reconstruction, trained with Adam (learning rate 8.42e-5, weight decay 1e-6, dropout 0.1).3
  • Model size is compact and subject-adaptive, ranging from roughly 550k to 790k parameters as the feature-reduction layer adjusts to each subject’s electrode configuration, supporting practical deployment for ECoG-based neuroprosthetic and neurorobotic control.2 3
  • For context, competing 2025 ECoG finger decoders report lower or comparable accuracy on the same data: the FingerFlex convolutional encoder–decoder reached up to 0.74 correlation (mean ~0.67 across subjects), and the newer HiLoFuseNet high-gamma/low-frequency fusion model reported average correlations of 0.631 (BCI Competition IV) and 0.534 (Stanford).4 5

Footnotes

  1. https://news.google.com/rss/articles/CBMipwFBVV95cUxPUDdlcmJ5VWtwUkRyaFJzV25DdndpSklMU1hmS3gyVWR4eVVGX3hqUEJWVmRQTG5UV2d6Q0ktZTJUYjBkUEJvcVMzdzBKOHMzeS00Sl9ucl9JRkgyc3J1TmtrTlFWREdla2MtcC1ERlVfczZKb2F0Ukl6dHRuOU1hV05Hdl9zeVpnU21WbVdkRDA2MGdNQjRaWXZIZkU5RWU2cUFSV3R3TQ?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2025.1627819/full ↩ ↩2 ↩3 ↩4

  3. https://pmc.ncbi.nlm.nih.gov/articles/PMC12283792/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  4. https://export.arxiv.org/pdf/2211.01960v2.pdf ↩

  5. https://doi.org/10.36227/techrxiv.176592065.57408366/v1 ↩

Graph View