EPIMix/NeuroMix Branch¶

The EPIMix branch handles classification of EPIMix and NeuroMix multicontrast outputs, rapid brain MRI sequences that produce multiple distinct contrasts from a single ~1-2 minute acquisition.

Overview¶

EPIMix and NeuroMix represent a paradigm shift in brain MRI: instead of acquiring each contrast separately (10-30 min total), these sequences produce all five essential contrasts in a single, motion-robust scan.

Aspect	Description
Provenance	`EPIMix`
Classification Branch	`epimix`
Source Module	`backend/src/classification/branches/epimix.py`
Output Types	12 distinct types
Series per Acquisition	5-9 typically
Scan Time	55-75s (EPIMix), 80s-4min (NeuroMix)

The Physics: Multicontrast EPI¶

The Clinical Problem¶

Standard brain MRI requires multiple sequences:

Contrast	Typical Acquisition	Purpose
T1-FLAIR	3-5 min	Gray/white matter
T2-FLAIR	4-6 min	White matter lesions
T2-w	2-3 min	Edema, inflammation
DWI/ADC	2-3 min	Stroke, restricted diffusion
T2*-w	2-3 min	Hemorrhage, iron
Total	15-20 min

Problems: Patient discomfort, motion artifacts, expensive scanner time, delays in acute stroke.

The EPIMix Solution¶

EPIMix uses 11 modular sequence blocks that can be dynamically combined:

flowchart LR
    subgraph inv["Inversion Modules"]
        a["a) T1-FLAIR inversion"]
        b["b) T2-FLAIR inversion"]
    end

    subgraph prep["Preparation Modules"]
        c["c) SE excitation (90°)"]
        d["d) GRE excitation (low flip)"]
        e["e) SE with 180° refocus"]
    end

    subgraph diff["Diffusion Modules"]
        f["f) Stejskal-Tanner prep"]
        g["g) Partial-Fourier EPI readout"]
    end

    subgraph readout["Readout Modules"]
        h["h) Full-Fourier EPI readout"]
        i["i) FLEET calibration"]
    end

    subgraph padding["Padding Modules"]
        j["j) Wait for magnetization recovery"]
        k["k) TR padding"]
    end

Acquisition Timeline (~75 seconds)¶

Time	Contrast	Modules Used
0-17s	T1-FLAIR	Inversion (a) + SE-EPI (c+h)
17-45s	T2-w/DWI/ADC	SE-EPI (e+g) + Diffusion (f)
45-47s	T2*-w	GRE-EPI (d+h)
47-52s	FLEET cal.	Sequential GRE (i+h)
52-75s	T2-FLAIR	Inversion (b) + SE-EPI (e+g)

Why Single-Shot EPI?¶

Each slice is acquired in 30-50 ms, effectively "freezing" patient motion:

Readout	Time per Slice	Motion Robustness
Multishot FSE	1-3 s	Motion → ghosting
PROPELLER	2-4 s	Motion-robust but slower
Single-shot EPI	30-50 ms	Motion frozen

Trade-off: EPI suffers from geometric distortion near air-tissue interfaces (skull base, sinuses).

NeuroMix: The Evolution¶

NeuroMix addresses EPIMix's main limitation—geometric distortion—by replacing EPI readouts with SSFSE (Single-Shot FSE) for critical contrasts.

Key Improvements¶

Contrast	EPIMix Readout	NeuroMix Readout
T1-FLAIR	SE-EPI	SE-EPI (unchanged)
T2-FLAIR	SE-EPI	SSFSE (no distortion)
T2-w	SE-EPI (b=0)	SSFSE (no distortion)
isoDWI/ADC	SE-EPI	SE-EPI (unchanged)
T2*-w	GRE-EPI	GRE-EPI (unchanged)

NeuroMix Optional Contrasts¶

NeuroMix adds three optional high-resolution sequences:

Optional Contrast	Readout	Resolution	Time Added
T1w 3D-EPI	3D GRE-EPI	1.2×1.2×1.2 mm³ (isotropic)	22.5s
SWI 3D-EPI	3D GRE-EPI	0.8×0.8×2.0 mm³	28.3s
T2w FSE	Multishot FSE	0.6×0.6×4.0 mm³	44.7s

Scan time range: 1:20 (minimal) to 4:00 (all options + 20 averages)

Output Types¶

EPIMix Core (Both EPIMix and NeuroMix)¶

Output	Base	Construct	Modifiers	Technique	Description
T1-FLAIR	T1w	-	FLAIR	SE-EPI	CSF-nulled T1w (TI ~580-670ms)
isoDWI	DWI	`isoDWI`	-	SE-EPI	Averaged 3-direction DWI
ADC	DWI	`ADC`	-	SE-EPI	Apparent diffusion coefficient
*T2-w**	T2*w	-	-	GRE-EPI	Gradient echo (no 180° refocus)

EPIMix-Only (EPI readouts)¶

Output	Base	Construct	Modifiers	Technique	Description
T2-FLAIR (EPI)	T2w	-	FLAIR	SE-EPI	CSF-nulled T2w, EPI readout
T2-w (EPI)	T2w	-	-	SE-EPI	b=0 from DWI block

NeuroMix-Only (SSFSE/FSE readouts)¶

Output	Base	Construct	Modifiers	Technique	Description
T2-FLAIR (SSFSE)	T2w	-	FLAIR	HASTE	No EPI distortion
T2-w (SSFSE)	T2w	-	-	HASTE	1.0×1.0 mm² resolution
T2-w (FSE)	T2w	-	-	TSE	High-res multishot (0.6×0.6 mm²)

NeuroMix Optional (3D-EPI)¶

Output	Base	Construct	Modifiers	Technique	Description
T1w 3D-EPI	T1w	-	-	EPI	Isotropic resolution, reformattable
SWI 3D-EPI	SWI	`SWI`	-	EPI	High-res susceptibility-weighted

Detection Strategy¶

Primary Detection: Text Keywords (90% confidence)¶

The EPIMix branch primarily uses text_search_blob pattern matching:

flowchart TB
    start["EPIMix Branch Input"] --> q1{T1 + FLAIR?}
    q1 -->|Yes| t1flair["1. T1-FLAIR"]
    q1 -->|No| q2{T2 + FLAIR?}

    q2 -->|Yes| q2a{SSFSE/HASTE?}
    q2a -->|Yes| t2flairssfse["2. T2-FLAIR (SSFSE)"]
    q2a -->|No| t2flairepi["2. T2-FLAIR (EPI)"]
    q2 -->|No| q3{'swi' in text?}

    q3 -->|Yes| swi["3. SWI (NeuroMix)"]
    q3 -->|No| q4{ADC keyword?}

    q4 -->|Yes| adc["4. ADC"]
    q4 -->|No| q5{isoDWI keyword?}

    q5 -->|Yes| isodwi["5. isoDWI"]
    q5 -->|No| q6{T2* indicators?}

    q6 -->|Yes| t2star["6. T2*-w"]
    q6 -->|No| q7{T1 + 3D-EPI?}

    q7 -->|Yes| t1epi["7. T1w 3D-EPI"]
    q7 -->|No| q8{T2 without FLAIR?}

    q8 -->|Yes| t2w["8. T2-w (EPI/SSFSE/FSE)"]
    q8 -->|No| fallback["Fallback"]

Secondary Detection: Physics Fallback (70-75% confidence)¶

When text keywords are unavailable, physics parameters distinguish outputs:

Output	TI Range	TE Range	Logic
T2-FLAIR	> 2000 ms	any	Long TI for CSF nulling
T1-FLAIR	500-750 ms	< 40 ms	Short TI, short TE
*T2-w**	0	< 60 ms	No TI, short TE (GRE)
T2-w	0	≥ 60 ms	No TI, long TE (ambiguous with DWI)

Database analysis (216 EPIMix stacks): - T2-FLAIR: TI 2461-3119ms (avg 2753) - T1-FLAIR: TI 548-671ms (avg 604) - T2*-w: TI=0, TE 16-51ms (avg 34) - T2-w/ADC/isoDWI: TI=0, TE 35-120ms (avg 93)

NeuroMix vs EPIMix Detection¶

def _is_neuromix(text_blob: str) -> bool:
    """NeuroMix-specific keywords: neuromix, mix2, mix3"""
    return any(kw in text_blob for kw in ["neuromix", "mix2", "mix3"])

def _has_ssfse_readout(text_blob: str) -> bool:
    """SSFSE/HASTE indicates NeuroMix readout for T2-FLAIR/T2w"""
    return any(kw in text_blob for kw in [
        "ssfse", "ss-fse", "haste",
        "single shot fse", "single-shot fse"
    ])

Technique Mapping¶

Each output maps to a specific technique from the technique-detection system:

Output Type	Technique	Technique Family
T1-FLAIR, T2-FLAIR (EPI), T2-w (EPI), isoDWI, ADC	SE-EPI	MIXED
T2*-w	GRE-EPI	MIXED
T2-FLAIR (SSFSE), T2-w (SSFSE)	HASTE	TSE (SS-TSE)
T2-w (FSE)	TSE	TSE
T1w 3D-EPI, SWI 3D-EPI	EPI	EPI

Clinical Significance¶

Why EPIMix Matters¶

Application	EPIMix Advantage
Acute stroke	Complete brain assessment in 1 min
Pediatric imaging	Reduced sedation need (short scan)
Uncooperative patients	Motion-robust single-shot acquisition
Screening	Cost-effective, high throughput
Emergency	CT-like door-to-scan time with MRI contrast

EPIMix vs Conventional Imaging¶

Aspect	EPIMix	Conventional
Scan time	55-75 s	15-20 min
Motion robustness	Excellent	Variable
Resolution	1.3×1.3 mm²	0.5-1.0 mm²
Geometric distortion	Present (EPI)	Minimal
SNR	Lower	Higher

NeuroMix vs EPIMix¶

Aspect	EPIMix	NeuroMix
T2-FLAIR distortion	Present	Eliminated (SSFSE)
T2-w distortion	Present	Eliminated (SSFSE)
Scan time (minimal)	72 s	80 s
Optional 3D contrasts	No	Yes (T1w, SWI)
Standalone capability	Limited	Suitable

Correct Classification Importance¶

Without EPIMix Branch	With EPIMix Branch
T1-FLAIR → "T1-weighted"	T1-FLAIR → base=T1w, modifiers=[FLAIR], technique=SE-EPI
T2-FLAIR → "FLAIR" (generic)	T2-FLAIR → base=T2w, modifiers=[FLAIR], technique=SE-EPI or HASTE
isoDWI → "DWI" (generic)	isoDWI → base=DWI, construct=isoDWI
T2*-w → confused with T2-w	T2-w → base=T2w, technique=GRE-EPI

Configuration Reference¶

Text Pattern Detection¶

Pattern	Output Type
`t1` + `flair`	T1-FLAIR
`t2` + `flair` (no `t1`) + `ssfse`/`haste`	T2-FLAIR (SSFSE)
`t2` + `flair` (no `t1`)	T2-FLAIR (EPI)
`swi`	SWI 3D-EPI
`adc`	ADC
`iso dwi` or `isodwi`	isoDWI
`t2star` or `t2*` or `gradient echo`	T2*-w
`t1` + `3d-epi`/`3depi`	T1w 3D-EPI
`t2w` + `ssfse`	T2w (SSFSE)
`t2w` + `fse` (not SSFSE)	T2w (FSE)
`t2w` (no FLAIR, no SSFSE)	T2w (EPI)

Output Type Mapping¶

From backend/src/classification/branches/common.py:

EPIMIX_OUTPUT_TYPES = {
    # Shared (EPIMix + NeuroMix)
    "t1_flair": {
        "base": "T1w",
        "construct": "",
        "modifiers": ["FLAIR"],
        "technique": "SE-EPI",
        "directory_type": "anat",
    },
    "t2star": {
        "base": "T2*w",
        "construct": "",
        "modifiers": None,
        "technique": "GRE-EPI",
        "directory_type": "anat",
    },
    "iso_dwi": {
        "base": "DWI",
        "construct": "isoDWI",
        "modifiers": None,
        "technique": "SE-EPI",
        "directory_type": "dwi",
    },
    "adc": {
        "base": "DWI",
        "construct": "ADC",
        "modifiers": None,
        "technique": "SE-EPI",
        "directory_type": "dwi",
    },

    # EPIMix-only (EPI readouts)
    "t2_flair_epi": {
        "base": "T2w",
        "modifiers": ["FLAIR"],
        "technique": "SE-EPI",
    },
    "t2w_epi": {
        "base": "T2w",
        "technique": "SE-EPI",
    },

    # NeuroMix-only (SSFSE readouts)
    "t2_flair_ssfse": {
        "base": "T2w",
        "modifiers": ["FLAIR"],
        "technique": "HASTE",  # SSFSE = HASTE
    },
    "t2w_ssfse": {
        "base": "T2w",
        "technique": "HASTE",
    },
    "t2w_fse": {
        "base": "T2w",
        "technique": "TSE",
    },

    # NeuroMix optional (3D-EPI)
    "t1w_3depi": {
        "base": "T1w",
        "technique": "EPI",
    },
    "swi_3depi": {
        "base": "SWI",
        "construct": "SWI",
        "technique": "EPI",
    },
}

Examples¶

Example 1: EPIMix T1-FLAIR¶

DICOM Fields:

SeriesDescription: EPIMix T1 FLAIR
ScanningSequence: IR\SE\EP
InversionTime: 580
EchoTime: 17.9

Classification:

base = "T1w"
construct = ""
modifiers = ["FLAIR"]
technique = "SE-EPI"
directory_type = "anat"

Example 2: EPIMix T2-FLAIR¶

DICOM Fields:

SeriesDescription: EPIMix T2 FLAIR
ScanningSequence: IR\SE\EP
InversionTime: 2750
EchoTime: 120

Classification:

base = "T2w"
construct = ""
modifiers = ["FLAIR"]
technique = "SE-EPI"
directory_type = "anat"

Example 3: NeuroMix T2-FLAIR (SSFSE)¶

DICOM Fields:

SeriesDescription: NeuroMix T2-FLAIR SSFSE
ScanningSequence: IR\SE
InversionTime: 2900
EchoTime: 108.8

Classification:

base = "T2w"
construct = ""
modifiers = ["FLAIR"]
technique = "HASTE"       # SSFSE → HASTE in technique taxonomy
directory_type = "anat"

Note: No EPI distortion compared to EPIMix T2-FLAIR.

Example 4: EPIMix isoDWI¶

DICOM Fields:

SeriesDescription: EPIMix isoDWI
ScanningSequence: SE\EP
bValue: 1000

Classification:

base = "DWI"
construct = "isoDWI"
modifiers = None
technique = "SE-EPI"
directory_type = "dwi"

Example 5: EPIMix T2*-w¶

DICOM Fields:

SeriesDescription: EPIMix T2*
ScanningSequence: GR\EP
EchoTime: 30

Classification:

base = "T2*w"
construct = ""
modifiers = None
technique = "GRE-EPI"
directory_type = "anat"

Note: GRE (gradient echo) readout, no 180° refocusing pulse.

Example 6: NeuroMix SWI 3D-EPI¶

DICOM Fields:

SeriesDescription: NeuroMix SWI 3D-EPI
ScanningSequence: GR\EP
ImageType: ORIGINAL\PRIMARY\M\SWI

Classification:

base = "SWI"
construct = "SWI"
modifiers = None
technique = "EPI"
directory_type = "anat"

Troubleshooting¶

Common Issues¶

Issue: T1-FLAIR and T2-FLAIR confused - Cause: Both contain "FLAIR" keyword - Solution: TI range distinguishes them (T1: 500-750ms, T2: >2000ms)

Issue: T2-w classified as DWI - Cause: Both from same SE-EPI block, identical physics - Solution: Text keyword detection preferred; physics fallback is ambiguous

Issue: NeuroMix SSFSE detected as EPI - Cause: Missing SSFSE/HASTE keyword - Solution: Check text_search_blob for SSFSE indicators

Issue: SWI 3D-EPI routed to SWI branch - Cause: Provenance detected as SWIRecon instead of EPIMix - Solution: EPIMix provenance should take precedence for NeuroMix outputs

Issue: Physics fallback gives low confidence - Cause: Text keywords unavailable, TI/TE values ambiguous - Solution: This is expected; physics detection has 70-75% confidence

Research Context¶

EPIMix and NeuroMix were developed at Karolinska Institutet, Stockholm, Sweden:

EPIMix (2018): Skare et al., "A 1-minute full brain MR exam using a multicontrast EPI sequence." Magn Reson Med 79:3045-3054.
NeuroMix (2022): Sprenger et al., "NeuroMix—A single-scan brain exam." Magn Reson Med 87:2178-2193.

Key contributors: Stefan Skare, Tim Sprenger, Ola Norbeck, Henric Rydén, Enrico Avventi.

EPIMix/NeuroMix Branch¶

Overview¶

The Physics: Multicontrast EPI¶

The Clinical Problem¶

The EPIMix Solution¶

Acquisition Timeline (~75 seconds)¶

Why Single-Shot EPI?¶

NeuroMix: The Evolution¶

Key Improvements¶

NeuroMix Optional Contrasts¶

Output Types¶

EPIMix Core (Both EPIMix and NeuroMix)¶

EPIMix-Only (EPI readouts)¶

NeuroMix-Only (SSFSE/FSE readouts)¶

NeuroMix Optional (3D-EPI)¶

Detection Strategy¶

Primary Detection: Text Keywords (90% confidence)¶

Secondary Detection: Physics Fallback (70-75% confidence)¶

NeuroMix vs EPIMix Detection¶

Technique Mapping¶

Clinical Significance¶

Why EPIMix Matters¶

EPIMix vs Conventional Imaging¶

NeuroMix vs EPIMix¶

Correct Classification Importance¶

Configuration Reference¶

Text Pattern Detection¶

Output Type Mapping¶

Examples¶

Example 1: EPIMix T1-FLAIR¶

Example 2: EPIMix T2-FLAIR¶

Example 3: NeuroMix T2-FLAIR (SSFSE)¶

Example 4: EPIMix isoDWI¶

Example 5: EPIMix T2*-w¶

Example 6: NeuroMix SWI 3D-EPI¶

Troubleshooting¶

Common Issues¶

Research Context¶

See Also¶