Annotation of skin image using learned feature representation

This patent covers a dual-network approach to automatically annotating dermatoscopic (skin lesion) images. In plain English, claim 1 describes a method that: receives a set of dermatoscopic images, each containing a region of lesion skin and a region of normal skin; trains a first convolutional neural network on the interior of the lesion region; trains a second convolutional neural network on the boundary between lesion and normal skin; then acquires a new dermatoscopic image and uses both trained CNNs together to identify the lesion region within it. The defining technical limitation is the explicit separation of two learned models — one specialized for lesion interior texture and one for the lesion/normal-skin border — combined at inference time. Dependent claims add preprocessing and image-conversion steps to the interior-network training pipeline. The architecture's premise is that interior appearance and boundary morphology carry complementary signal for accurate segmentation.

The patent is assigned to International Business Machines Corporation, with four named inventors (Mani Abedini, Rajib Chakravorty, Rahil Garnavi, Munawar Hayat). It was granted in 2018. The family is small — 2 members in a single country — indicating focused, US-centric prosecution rather than aggressive international filing. Forward citations stand at 5, which is modest. For a granted CNN/medical-imaging patent from this era, a low forward-citation count more likely reflects that it sits inside a tightly clustered IBM dermatology-AI portfolio (where related IBM filings cross-reference each other) than that the work is unimportant; citation-weighted ranking tends to under-reward such defensive, portfolio-internal filings.

Where this patent stands: This is best read as one node in a coordinated IBM cluster of skin-lesion deep-learning patents filed in the 2017–2019 window — alongside US-10223788-B2, US-10373312-B2, and US-10176574-B2, all IBM, all on lesion segmentation via deep networks. Rather than a foundational anchor, it functions as a specific architectural refinement (interior network + boundary network) within that portfolio. It does not appear to underpin a single flagship product so much as protect a method family around automated dermatology imaging, an area where IBM (via its Watson Health imaging research) was active. Relative to the field's heavyweights, its citation footprint is small; its value lies in portfolio breadth and the specificity of the two-network claim, not in being a widely-cited reference work.

The surrounding space is moderately crowded and clearly clustered around medical-image deep learning, but it is fragmented across many single-patent assignees rather than dominated by one player. The citation chart is led by Electro Optical Sciences Inc (582 and 531 citations across two near-duplicate entries), 12 Sigma Technologies (2 patents, 413 citations), and Sharp Laboratories of America (302), with Eyenuk, Arterys, and The Procter & Gamble Company also ranking. Notably, IBM is the most relevant assignee for the subject patent's immediate neighborhood — it holds the four closest patents ([1]–[4]) — yet IBM does not appear in the citation-ranked TOP_ASSIGNEES list at all, confirming that the chart ranks by citation impact rather than topical proximity. The field's clustering is unambiguous in the technology mix: Computer Vision (G06V) runs at 34.8× the corpus baseline, Healthcare IT (G16H) at 29.3×, and Image Processing (G06T) at 27.7× — a tightly concentrated medical-vision space.

The closest related patents:

• US-10223788-B2 — Skin lesion segmentation using deep convolution networks guided by local unsupervised learning (IBM · 17 citations · 4-member family). The same IBM team's adjacent filing; it pursues lesion segmentation with CNNs but layers in local unsupervised guidance, a direct sibling to the subject patent's interior/boundary split.
• US-10373312-B2 — Automated skin lesion segmentation using deep side layers (IBM · 5 citations · 2-member family). Another IBM companion patent attacking the same segmentation problem with a different architectural device (deep side layers), reinforcing the portfolio character.
• US-10176574-B2 — Structure-preserving composite model for skin lesion segmentation (IBM · 3 citations · 4-member family). Completes the IBM cluster; a composite-model variant of lesion segmentation.
• US-7689016-B2 — Automatic detection of critical dermoscopy features for malignant melanoma diagnosis (Stoecker & Associates · 71 citations · 2-member family). A foundational pre-deep-learning dermoscopy reference and the most-cited neighbor in this set; relevant as prior conceptual groundwork for automated lesion feature detection.
• US-11875479-B2 — Fusion of deep learning and handcrafted techniques in dermoscopy image analysis (Nabin K Mishra · 0 citations · 2-member family). A 2024 on-point filing combining learned and handcrafted features; as a recent, lightly-cited but directly relevant patent, this is strategic IP the citation-weighted score under-rewards.

The non-IBM neighbors split between dedicated dermatology players (Stoecker & Associates, Nabin K Mishra, Bostel, Lululab, L'Oréal) and general medical-CNN heavyweights whose high citations dominate the chart — **12 Sigma Technologies' US-10304