The compression of medical images demands both aggressive bitrate reduction and the faithful reconstruction of diagnostically critical structures. Standard codecs allocate bits uniformly, ignoring clinical salience, while region-of-interest methods require external segmentation annotations. We introduce the Sparsity-Controlled Anatomical Attention (SCAA) framework: a self-supervised mechanism that learns to prioritise anatomically significant regions during compression without manual labels. SCAA couples a SparsityPriorGenerator (producing soft organ-attention maps via a learnable temperature) with a linear-complexity AnatomicalAttention block that conditions latent feature routing, achieving <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\mathcal {O}(Ld^{2})$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi mathvariant="script">O</mi> <mo stretchy="false">(</mo> <mi>L</mi> <msup> <mi>d</mi> <mn>2</mn> </msup> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation> complexity – about three orders of magnitude lower than standard self-attention. Evaluated on a dedicated 2D CT dataset derived from LUNA16 (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$6\,216$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>6</mn> <mspace width="0.166667em" /> <mn>216</mn> </mrow> </math></EquationSource> </InlineEquation> axial slices, patient-level split), our full compression pipeline attains 29.84 dB PSNR at 0.310 bpp (SSIM 0.9695, LPIPS 0.052) – a compression ratio exceeding <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$50\times $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>50</mn> <mo>×</mo> </mrow> </math></EquationSource> </InlineEquation> – while the ROI PSNR remains within 0.8 dB of the global PSNR and the spatial bitrate–organ correlation reaches 0.188. At equivalent low bitrates, SCAA substantially outperforms both JPEG2000 and state-of-the-art learned codecs. These results demonstrate that SCAA preserves diagnostic image quality under extreme compression, offering a computationally efficient, annotation-free alternative to supervision-dependent ROI coding. Our code is publicly available at <a href="https://github.com/BouraiNour/SCAA">https://github.com/BouraiNour/SCAA</a>.

Preserving Diagnosis, Reducing Bits: Sparsity-Controlled Linear Anatomical Attention for Medical Image Compression

期刊论文
发表日期：

2026年8月5日

查看全文

Nour El Houda Bourai,
Hayet Farida Merouani,
Akila Djebbar

摘要

The compression of medical images demands both aggressive bitrate reduction and the faithful reconstruction of diagnostically critical structures. Standard codecs allocate bits uniformly, ignoring clinical salience, while region-of-interest methods require external segmentation annotations. We introduce the Sparsity-Controlled Anatomical Attention (SCAA) framework: a self-supervised mechanism that learns to prioritise anatomically significant regions during compression without manual labels. SCAA couples a SparsityPriorGenerator (producing soft organ-attention maps via a learnable temperature) with a linear-complexity AnatomicalAttention block that conditions latent feature routing, achieving $\mathcal {O}(Ld^{2})$ $O (L d^{2})$ complexity – about three orders of magnitude lower than standard self-attention. Evaluated on a dedicated 2D CT dataset derived from LUNA16 ( $6\,216$ $6 216$ axial slices, patient-level split), our full compression pipeline attains 29.84 dB PSNR at 0.310 bpp (SSIM 0.9695, LPIPS 0.052) – a compression ratio exceeding $50\times $ $50 \times$ – while the ROI PSNR remains within 0.8 dB of the global PSNR and the spatial bitrate–organ correlation reaches 0.188. At equivalent low bitrates, SCAA substantially outperforms both JPEG2000 and state-of-the-art learned codecs. These results demonstrate that SCAA preserves diagnostic image quality under extreme compression, offering a computationally efficient, annotation-free alternative to supervision-dependent ROI coding. Our code is publicly available at https://github.com/BouraiNour/SCAA.