Existing hyperchaotic medical image encryption schemes derive all chaotic initial conditions solely from statistical features of deep-learning-segmented anatomical regions, creating a structural chosen-plaintext-attack (CPA) vulnerability: images with similar segmentation statistics yield cipher images from the same keystream, resulting in a Cipher Cross Correlation (CCC) close to unity (experimentally measured at 0.9141). We introduce a semantically conditioned key schedule that binds deep-learning semantic features with cryptographic hash functions to initialise chaotic dynamical systems. As a concrete instantiation, U-Net segmentation statistics and a digest of the whole plaintext image are fused via a nonlinear $$\tanh$$ tanh coupling. The cryptographic guarantees structural CPA resistance, full-image plaintext sensitivity, and an effective key space of $$2^{256}$$ 2 256 derive from the keyed hash component under standard hash assumptions (complexity-theoretic IND-CPA is not claimed), while the U-Net component provides semantic key conditioning , enabling content-aware deployment scenarios such as audit traceability. The framework is drop-in compatible with existing permutation-diffusion pipelines and generalises beyond U-Net and to other semantic extractors and hash primitives. Experiments on six medical imaging datasets across five modalities confirm near-ideal entropy ( $$\approx 7.997$$ ≈ 7.997 bits), NPCR ( $$\approx 99.61\%$$ ≈ 99.61 % ), and UACI ( $$\approx 33.47\%$$ ≈ 33.47 % ), and a reduction of more than three orders of magnitude in adversarial CCC relative to the base architecture, at under 2 ms of additional computation per $$256\!\times \!256$$ 256 × 256 image.